chunk_uid
stringlengths
40
40
chunk_type
stringclasses
2 values
chunk_index
int64
0
6.71k
total_chunks
int64
1
6.71k
section_title
stringlengths
1
157
embed_text
stringlengths
1
83.3k
spans
dict
paper_doi
stringlengths
0
63
paper_id_arxiv
stringlengths
9
16
title
stringlengths
7
245
authors
listlengths
1
768
categories
listlengths
1
7
year
int64
2k
2.02k
language
stringclasses
2 values
discipline
stringclasses
8 values
dense_vector
listlengths
1.02k
1.02k
e044e4eacd87edd699e9623c98e94c22c71cc760
subsection
54
80
Times in the non-exponential regime
Denoting the eigenvalues and projectors of this Liouvillian by \tilde{\gamma }_j^{\text{st}} and \tilde{P}_j^{\text{st}}, with j=\text{st},0,\pm , we show in Appendix  that we get in analogy to (REF -) and (REF -)\tilde{\gamma }_{\text{st}}^{\text{st}} &= 0\quad ,\\ \tilde{\gamma }_0^{\text{st}} &= -i\Big (\Gamma ^{(1...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.03289768099784851, 0.04708824306726456, -0.024688521400094032, -0.049193937331438065, 0.013000383041799068, 0.011253265663981438, 0.055663611739873886, 0.057860858738422394, 0.031097160652279854, 0.05358843132853508, -0.02078230306506157, 0.03671234846115112, -0.0014982102438807487, -0....
92f1f3ecb4b2c01944663e18ab4480c42feee9b0
subsection
55
80
Times in the non-exponential regime
This goes beyond the scope of this paper. However, in Ref. divincenzoloss05, such an analysis has been performed in bare perturbation theory (i.e. using the unrenormalized tunneling) with the result (note that we slightly changed the result such that it is valid for a Lorentzian cutoff function in the bath){\langle \si...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.00913948379456997, 0.015410481952130795, -0.022612212225794792, 0.015326564200222492, 0.027723610401153564, -0.057827454060316086, 0.021925607696175575, 0.030134359374642372, -0.010573727078735828, 0.044858235865831375, -0.013930466026067734, -0.012526738457381725, 0.012152919545769691, ...
ca8678795d51071c685d65034381b512437b7606
subsection
56
80
Times in the non-exponential regime
There is a Z-factor renormalization Z=\tilde{\Delta }^2/\Delta ^2 for F^c_{x,y} and F^s_{x,y}, and terms \sim \Delta or \sim \Delta ^3 in the Bloch-Redfield solution are replaced by \sqrt{Z}\Delta =\tilde{\Delta }^2/\Delta and Z^2\Delta ^3=\tilde{\Delta }^4/\Delta , respectively.The most interesting correction in O(\al...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.03918934240937233, 0.03952507674694061, -0.04398118704557419, 0.0015937838470563293, 0.01223140861839056, -0.03735806420445442, 0.02308935858309269, 0.001960993278771639, 0.0053984541445970535, 0.017183488234877586, -0.021746423095464706, 0.02702660672366619, -0.035343658179044724, 0.02...
a66789570b2ad5511baac4be6fd97aca5e8ce852
subsection
57
80
Times in the non-exponential regime
In this case one can use the approximation H((z_+-\Omega _0)t)\approx -\gamma -\ln (-i(z_+-\Omega _0)t) and e^{-iz_+t}\approx e^{-i\Omega _0t} in (REF ) leading to&{i\over 2\pi }\int dE e^{-iEt}{1\over E-z_+}\ln {-i(E-\Omega _0)\over \Omega _0}\,\approx \\ & =\,-\Big (\gamma + \ln (\Omega _0 t)\Big )e^{-i\Omega _0 t}\...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1749, "openalex_id": "", "raw": "A.J. Leggett, S. Chakravarty, T.A. Dorsey, M.P.A. Fisher, A. Garg, and W. Zwerger, Rev. Mod. Phys. 59, 1 (1987).", "source_ref_id": "c23ab74cf76ec9d3bef89e71e0bd211e2c5eb7d3", "start": ...
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.040472060441970825, 0.05420692637562752, -0.031223921105265617, 0.020861729979515076, -0.011300739832222462, -0.00894292164593935, 0.022387826815247536, -0.0022109313867986202, 0.016726011410355568, 0.017290666699409485, -0.040655191987752914, -0.008515615016222, 0.0073595973663032055, ...
5846d1bb93f0609ec648f6800563e7583a087a48
subsection
58
80
Times in the non-exponential regime
In this case and for \Omega t\gg 1 our solution reduces to{\langle \sigma _{z}\rangle }(t)\,\approx \, (1+\alpha )\cos (\tilde{\Delta }t)e^{-{\Gamma \over 2}t} - 2\alpha {1\over (\tilde{\Delta } t)^2}e^{-\Gamma t}\quad ,which, up to the missing exponential for the second term on the r.h.s., agrees with the NIBA result ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 455, "openalex_id": "", "raw": "A.J. Leggett, S. Chakravarty, T.A. Dorsey, M.P.A. Fisher, A. Garg, and W. Zwerger, Rev. Mod. Phys. 59, 1 (1987).", "source_ref_id": "c23ab74cf76ec9d3bef89e71e0bd211e2c5eb7d3", "start": 0...
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.02208895981311798, 0.017482008785009384, -0.024651767686009407, 0.0008585595642216504, -0.012402157299220562, -0.05421558395028114, 0.02608571946620941, 0.018778666853904724, 0.028923112899065018, 0.019541407003998756, -0.010068172588944435, -0.001194642041809857, -0.031302861869335175, ...
2d6b04c8ec74b49629a55f2921054e897039abfa
subsection
59
80
Times in the non-exponential regime
Keeping only the consistent terms falling off \sim \alpha /(\Omega t) we obtain for large times \Omega t\gg 1:F_x^0(t) \,&=\, -{\langle \sigma _x\rangle }_{\text{st}} \,-\, \Big (1+2\alpha {\tilde{\Delta }^2\over \Omega ^2}\Big ){\tilde{\Delta }^2\over \Delta \Omega } {\langle \tilde{\sigma }_z\rangle }_0 \,+\, \pi \al...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.028830986469984055, 0.04807708039879799, -0.022741222754120827, 0.0030773147009313107, 0.02722841687500477, -0.05863877758383751, 0.006513299886137247, 0.014278129674494267, 0.02124549075961113, 0.0239774901419878, -0.0009710807353258133, -0.002844560658559203, -0.04191100597381592, 0.0...
38ae904dc8c6551b046e7dcc4fee5890fd0f65e0
subsection
60
80
Times in the non-exponential regime
As we discuss later on, the main result is that the function f_t has to be replaced by the power-lawf_t\,\rightarrow \,\left({1\over \Omega t}\right)^{2\alpha {\epsilon ^2\over \Omega ^2}} \Big (1-2\alpha \gamma {\epsilon ^2\over \Omega ^2}+\alpha {\tilde{\Delta }^2\over \Omega ^2}\Big )\quad .with a power-law exponent...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.055910781025886536, 0.03772146627306938, -0.04657197371125221, -0.028062207624316216, -0.0005817681667394936, -0.03872859477996826, -0.005352267064154148, 0.031281959265470505, 0.03570721298456192, -0.021363290026783943, -0.004352770280092955, 0.002935545053333044, -0.01620558090507984, ...
ab1c03698aa6aedbafc47b1e3058f79d51956629
subsection
61
80
Times in the non-exponential regime
The logarithmic terms are a result of a combination of logarithmic terms arising from the terms \sim \alpha \ln (\Omega t) appearing explicitly in (-) and those arising from the functions H_t and \tilde{H}_t, which, for small argument, can be expanded asH^\prime _t\,&=\,-\gamma \,-\,\ln (\Omega t)\,+\,O(\Omega t)\quad ...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.027361081913113594, 0.04352136328816414, -0.04016417637467384, -0.011277099139988422, 0.019425909966230392, -0.013299042358994484, 0.006813565269112587, 0.012032466940581799, 0.028948115184903145, -0.017671016976237297, -0.004574168473482132, 0.031801726669073105, -0.04022521525621414, ...
51afb0e864b2ccd38f7614abf92b377350c47ede
subsection
62
80
Times in the non-exponential regime
In this regime our full solution (REF -) is needed to calculate all terms one order beyond Bloch-Redfield. In this case the time dependence of the preexponential functions is governed by a complicated combination of slowly varying logarithmic terms and terms arising from the functions H_t and \tilde{H}_t containing the...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.01950627751648426, 0.006769197527319193, -0.016636809334158897, 0.02463468909263611, -0.0008084688452072442, -0.022284166887402534, 0.02539784647524357, 0.0340978279709816, 0.009310508146882057, 0.0010178599040955305, -0.020589960739016533, 0.005323016084730625, 0.01958259381353855, 0.0...
72fc566d2e7bc1e6485632debfe9b3156b05b8fe
subsection
63
80
Real-time renormalization group
In this section we will present the real-time renormalization group approach to calculate the Liouvillian L(E) beyond perturbation theory by including the leading logarithmic series at low and high energies. This provides the basis for the renormalized perturbation theory in the non-exponential regime together with the...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.032296665012836456, 0.036967165768146515, 0.009196002036333084, 0.014652551151812077, -0.00655930582433939, -0.017873059958219528, 0.026878273114562035, -0.013126243837177753, -0.010913098230957985, 0.0036287957336753607, -0.036967165768146515, 0.015148600563406944, -0.009287580847740173,...
27954621a67c1f015fdf3902a57f496fde8f01e5
subsection
64
80
RG equations
The leading order RG equations to determine the Liouvillian L(E) for the ohmic spin boson model have been derived in Ref. kashubaschoeller13. Using the definitions (REF ), (REF ) and (REF ), they read{d\over dE}\tilde{L}_\Delta (E)\,&=\,2\alpha \sum _i Z^{\prime }(E) G(E) P_i(E) Z^{\prime }(E) G(E) {\tilde{L}_\Delta (E...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.019319603219628334, 0.0260341577231884, -0.0031226491555571556, -0.02627832256257534, 0.012513486668467522, -0.0063673811964690685, 0.0331454798579216, -0.012475335970520973, 0.017595183104276657, 0.020235223695635796, -0.0355566143989563, 0.009614020586013794, 0.020708294585347176, 0.0...
8a379ad5ae60cfd315bdf5d5c302ce0332e9feaa
subsection
65
80
RG equations
For \omega =0 and \Lambda \rightarrow 0, we obtain the Liouvillian L(0^+) from which we can obtain the stationary state \rho _{\text{st}} via (REF ). For \omega =\pm \eta (\eta =0^+) and \Lambda < -\Gamma , we obtain a jump between L(-i\Lambda +\eta ) and L(-i\Lambda -\eta ) indicating the branch cut starting at z_0=-i...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.04978080466389656, 0.03720588609576225, -0.006462958641350269, 0.014970866963267326, 0.020235851407051086, -0.03238346800208092, 0.027255829423666, 0.010293425060808659, 0.03284129127860069, 0.0281562190502882, -0.035191457718610764, 0.0004516249755397439, -0.009507493115961552, 0.01109...
8e9c838d8c16a809ea4d8a7499fda3b4228f7046
subsection
66
80
RG equations
Therefore, the RG equations provide a well-defined set of differential equations which can be systematically truncated at order \alpha , including secular terms and the leading logarithmic series to all orders.We note that the vertex renormalization is very essential. At large energies we show in Section REF that the v...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.017579324543476105, 0.04391779378056526, 0.0036261172499507666, -0.014512098394334316, -0.004677137825638056, -0.06250426918268204, 0.027803411707282066, 0.012650399468839169, 0.0035383731592446566, 0.00849209539592266, -0.036013200879096985, -0.0124215018004179, -0.024736184626817703, ...
18240f26f08f08f3d48ec00a80eb6ae433f5a2bd
subsection
67
80
Large energies
For large energies, where \Omega \ll |E| \ll D, we will show in this section that \tilde{L}_\Delta (E) and Z^{\prime }(E) are given by (REF -). For |E|\gg |\lambda _i(E)|, the RG equations (REF -) can be approximated by{d\over dE}\tilde{L}_\Delta (E)\,&=\,2\alpha \sum _i Z^{\prime }(E) G(E) P_i(E) Z^{\prime }(E) G(E) {...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.021957306191325188, 0.05075052008032799, -0.007682768162339926, -0.0558469332754612, -0.024368183687329292, -0.030288569629192352, 0.061431627720594406, 0.01730339415371418, -0.002262105932459235, 0.013259832747280598, -0.021575838327407837, 0.010322527959942818, -0.020004188641905785, ...
750effa1610eb2dc854e7be7766a4c65f1930ed9
subsection
68
80
Large energies
Furthermore we find{d\over dE} Z(E)^2 g(E)^2\,&=\,0\quad ,\\ {d\over dE} Z(E)\,&=\,2\alpha Z(E) {1\over E}\quad ,with the solutionZ(E)^2 g(E)^2\,&=\, 1\quad ,\\ Z(E)\,&=\,\left({-iE\over D}\right)^{2\alpha }\quad .In conclusion, (REF ), () and () prove the form (REF -) of the Liouvillian at large energies which was u...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.02840128168463707, 0.03281179815530777, -0.016390638425946236, -0.030949918553233147, -0.007168996147811413, -0.017336837947368622, 0.057290926575660706, 0.00264974357560277, -0.0529261939227581, 0.0016501282807439566, 0.008668418973684311, 0.009507791139185429, 0.0018084641778841615, 0...
871bbb8bbfb646ba4e9a42e21072bdf8a605046c
subsection
69
80
The non-exponential regime
In the non-exponential regime (REF ), where \alpha \ln (-i(E-\lambda _i(E))/\Omega )\ll 1, the RG equations can be solved perturbatively around the solution (REF -) at high energies evaluated at E=i\Omega , see Ref. schoeller09 for details. Denoting the latter by\tilde{L}_0\,&=\, \left(\begin{array}{cc} 0 & \Delta \tau...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.04517362639307976, 0.012712713330984116, -0.0023750492837280035, -0.0077260639518499374, 0.01078978180885315, -0.022510506212711334, 0.03062955103814602, 0.028966061770915985, 0.02832508459687233, 0.009172077290713787, -0.042304493486881256, 0.00761541910469532, -0.023517757654190063, 0...
0914c6ae16169d7b5c7042318e2925cb17bb8418
subsection
70
80
The non-exponential regime
For large |E|\gg |\lambda _i(E)| but not exponentially large the solution (REF -) is consistent with the result at large energies, given by (REF ), () and (), when expanded in \alpha \ln (-iE/\Omega ). As a consequence it is straightforward to see that (REF -) is indeed the solution of the RG equations in the non-expon...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.03982026129961014, 0.05141543224453926, -0.00839124247431755, -0.04146799445152283, -0.013433616608381271, -0.021634148433804512, 0.04824201762676239, 0.024639740586280823, 0.017438527196645737, -0.010687392204999924, -0.004142222460359335, 0.04509911313652992, -0.010084748268127441, 0....
15dd7ab2532a4bcebe80179ed21e5a36e09d9bd3
subsection
71
80
The non-exponential regime
After a straightforward calculation we obtainA\tilde{L}_\Delta (E)A&= \Omega \left(\begin{array}{cc} 0 & 0 \\ 0 & \sigma _z \end{array}\right)+ 2\alpha {\tilde{\Delta }^2\over \Omega }{\cal {L}}_0(E) \left(\begin{array}{cc} 0 & 0 \\ 0 & \tau _-\sigma _z \end{array}\right)\\ & \hspace{-28.45274pt} -\alpha {\tilde{\Delt...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.021101640537381172, 0.06579195708036423, -0.02746417373418808, -0.02708272635936737, 0.005828507710248232, 0.0030534821562469006, 0.007716669701039791, 0.0283186137676239, 0.015357050113379955, 0.026060448959469795, 0.014845911413431168, 0.026014676317572594, 0.0022524436935782433, -0.0...
042de0796dafde19ffb0475991ca0fb996f8dcaa
subsection
72
80
Exponentially large times
For exponentially large times, where higher powers in \alpha \ln (\Omega t) become significant and can no longer be treated in lowest order to analyze the corrections to Bloch-Redfield, we need a solution of the RG equations exponentially close to the branching points z_i. Analytically, such an analysis is very complic...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.046106915920972824, 0.02451813779771328, -0.03588466718792915, 0.0032783665228635073, -0.00800997018814087, -0.05321671813726425, 0.04751056805253029, -0.014143319800496101, 0.01995627023279667, 0.00965773593634367, -0.02016986906528473, 0.0028607037384063005, -0.017255764454603195, 0.0...
306b7a627c3d11b05dd263eb9d66f6711b6617c8
subsection
73
80
Exponentially large times
As already discussed in Ref. kashubaschoeller13, we note that there is no change of the power-law exponent of the 1/t^2 parts, in particular for the time dynamics of {\langle \sigma _z\rangle }(t), see (REF ), in contrast to the NIBA solution which predicts an incorrect power-law exponent 2-2\alpha , .As discussed in d...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 304, "openalex_id": "", "raw": "A.J. Leggett, S. Chakravarty, T.A. Dorsey, M.P.A. Fisher, A. Garg, and W. Zwerger, Rev. Mod. Phys. 59, 1 (1987).", "source_ref_id": "c23ab74cf76ec9d3bef89e71e0bd211e2c5eb7d3", "start": 0...
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.0665772333741188, 0.0005625654594041407, -0.050100743770599365, 0.005396814085543156, -0.010152876377105713, -0.0299475509673357, 0.01717826910316944, 0.02901693433523178, 0.023662075400352478, -0.02102278359234333, -0.0020729105453938246, -0.00027556170243769884, -0.03151892125606537, ...
6014e70145191c547eac3b986c714347bfc401e2
subsection
74
80
Summary
In this work we have presented the solution for the time dynamics of the ohmic spin boson model at finite bias by systematically expanding one order beyond Boch-Redfield. Using real-time RG and perturbation theory we have set up a renormalized perturbation theory to study analytically the whole time regime from exponen...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.0281695444136858, 0.0010672304779291153, 0.0005402914248406887, -0.0033514280803501606, -0.004726715385913849, -0.05591181665658951, 0.010590284131467342, -0.015824386849999428, 0.004265107214450836, 0.030122796073555946, -0.03415137529373169, 0.025392265990376472, -0.0002012382901739329,...
3ecb7db322a30909e9a583536031efdbc60222ae
subsection
75
80
Summary
We solved this problem by expanding all analytic parts of \Sigma (E) around E=z_i and keeping \Sigma (z_i) in the denominator whereas all other higher terms of the Taylor expansion are at least of O(\alpha ^2) and can be taken as a small correction. The non-analytic terms of \Sigma (E) are more subtle and are some func...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 854, "openalex_id": "", "raw": "H. Schoeller and F. Reininghaus, Phys. Rev. B 80, 045117 (2009); ibid. Phys. Rev. B 80, 209901(E) (2009).", "source_ref_id": "a365318024c850d29abd3ad05207e9287c8dd63a", "start": 612 ...
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.01741327904164791, -0.011224773712456226, -0.0365663580596447, 0.021274417638778687, 0.025517093017697334, -0.07319376617670059, 0.0015270956791937351, -0.005387281067669392, -0.024418270215392113, 0.048409219831228256, -0.010553271509706974, 0.04206047207117081, -0.043220337480306625, ...
3f01544d1019acda01b32d06c13933c4897df265
subsection
76
80
Summary
Therefore, second order terms are needed for the Liouvillian to calculate the stationary state and all terms of the time evolution of the purely decaying modes up to first order in \alpha . Again this problem occurs only for times of the order of the inverse decay rate, since for small times |E-L_0|\sim 1/t is much lar...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1954, "openalex_id": "", "raw": "H. Schoeller, Eur. Phys. J. Spec. Top. 168, 179 (2009).", "source_ref_id": "0dae0422b0c362aee37eca69b21b203ccbc1c56b", "start": 1859 }, { "arxiv_id": "", "doi": "10....
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.005728909745812416, 0.031978148967027664, -0.013631906360387802, -0.00635062251240015, 0.015386432409286499, -0.02499055676162243, 0.0027366795111447573, 0.005892919842153788, 0.015172838233411312, 0.02810293436050415, -0.018201302736997604, -0.023342827335000038, -0.0044206432066857815, ...
b722b9bd561612595897fb40e0052998254d6e03
subsection
77
80
Projectors for
To calculate the projectors of the matrix \tilde{L}_0+\tilde{\Sigma }_a^i up to O(\alpha ), we first set up the matrix \tilde{\Sigma }_a^i=\tilde{\Sigma }_a(z_i) by setting E=z_i in () and use& {\cal {F}}_0(0)\,\sim \,O(\alpha )\quad ,\quad {\cal {F}}_0(z_0)\,=\,{\cal {F}}_\sigma (z_\sigma )\,=\,0 \quad ,\\ & {\cal {F}...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.026226475834846497, 0.022839460521936417, -0.0027347852010279894, 0.009154092520475388, 0.028438715264201164, 0.004630445037037134, -0.009123578667640686, 0.011885063722729683, 0.0356399342417717, 0.033320896327495575, -0.004607560113072395, 0.027263939380645752, -0.008520934730768204, ...
0f7328a9c7400295914466d8a1a40703ba673c59
subsection
78
80
Projectors for
Due to the matrix structure of \tilde{\Sigma }_a(E), one projector is exactly known (in all orders of perturbation theory, see () and (REF ))A \tilde{P}^i_{\text{st}} A\,&=\, \left(\begin{array}{cc} \tau _+ & 0 \\ 0 & 0 \end{array}\right)\quad .Using usual perturbation theory it is straightforward to calculate the proj...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ 0.009232919663190842, 0.004578307271003723, -0.009118461981415749, 0.0008703553467057645, 0.03586340695619583, -0.006146377883851528, 0.01721443608403206, -0.0046622431837022305, 0.045203156769275665, 0.0594874769449234, -0.01916784793138504, 0.002731723478063941, -0.01050721574574709, -0....
2bac19d99fde9e9e88212716271447f0676fde7d
subsection
79
80
Projectors for
This is a particular advantage for the spin boson model.To derive the formula () for the eigenvalue \tilde{\gamma }_0^\text{st} of \tilde{L}_a(0) up to second order in \alpha , we relate it to the eigenvalue \tilde{\gamma }_0^0=z_0=-i(\Gamma ^{(1)}+\Gamma ^{(2)}+O(\alpha ^3)) of \tilde{L}_a(z_0). We first note that, du...
{ "cite_spans": [] }
10.1103/PhysRevB.98.115425
1802.09846
Dissipative quantum mechanics beyond Bloch-Redfield: A consistent weak-coupling expansion of the ohmic spin boson model at arbitrary bias
[ "Carsten J. Lindner", "Herbert Schoeller" ]
[ "cond-mat.stat-mech", "cond-mat.mes-hall" ]
2,018
en
Physics
[ -0.023135095834732056, 0.039524998515844345, -0.015611611306667328, -0.00470790034160018, 0.007195380982011557, -0.006760453339666128, 0.034275345504283905, 0.01616099290549755, 0.04166148602962494, 0.02609565481543541, -0.015642132610082626, 0.036289747804403305, -0.03259667754173279, -0....
556c4a6e6a58ea1085bb757993996936ed01f19b
abstract
0
106
Abstract
We study the computational tractability of PAC reinforcement learning with rich observations. We present new provably sample-efficient algorithms for environments with deterministic hidden state dynamics and stochastic rich observations. These methods operate in an oracle model of computation -- accessing policy and va...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.026956412941217422, -0.002234926214441657, -0.006967935245484114, 0.00026744743809103966, -0.013806198723614216, -0.02793276309967041, 0.07395850867033005, 0.04991589114069939, 0.014507950283586979, 0.04607151448726654, -0.018215028569102287, -0.03951166197657585, 0.0008452482870779932, ...
cef249bc1dbbe1fcbff8183e0f920a6b74618d39
subsection
1
106
Introduction
We study episodic reinforcement learning (RL) in environments with realistically rich observations such as images or text, which we refer to broadly as contextual decision processes. We aim for methods that use function approximation in a provably effective manner to find the best possible policy through strategic expl...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1038/nature14236", "end": 498, "openalex_id": "https://openalex.org/W2145339207", "raw": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjela...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.017729677259922028, 0.007300904020667076, -0.02813556417822838, 0.01573089323937893, -0.02441263571381569, 0.005263974890112877, 0.05770536884665489, 0.021742504090070724, 0.02572481520473957, 0.007972251623868942, 0.006805022247135639, -0.033384282141923904, -0.010733931325376034, 0.03...
3daebfc2ebf3f50c4bc52d53b9d63eaf85a1709d
subsection
2
106
Introduction
Together, these results advance our understanding of efficient reinforcement learning with rich observations.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.015393589623272419, 0.01135067455470562, -0.033563822507858276, -0.011159971356391907, -0.03231280669569969, -0.07463373243808746, 0.02346416376531124, 0.016431016847491264, 0.02271660603582859, 0.0006746137514710426, -0.048148829489946365, -0.014485840685665607, -0.04851498082280159, 0...
dfc6272e8e6d193a6b4ef95af85d178b0b9a1653
subsection
3
106
Related Work
There is abundant work on strategic exploration in the tabular setting , , , , , , , . The computation in these algorithms often involves planning in optimistic models and can be solved efficiently via dynamic programming. To extend the theory to the more practical settings of large state spaces, typical approaches inc...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 86, "openalex_id": "", "raw": "Michael Kearns and Satinder Singh. Near-optimal reinforcement learning in polynomial time. Machine Learning, 2002.", "source_ref_id": "7ba9b49b98b82a5addc01e94c324c957917e57f2", "start": ...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0593438483774662, -0.023765012621879578, -0.04105837270617485, 0.04438577964901924, -0.033548809587955475, 0.0035277383867651224, 0.04887320101261139, -0.007936934009194374, 0.021002348512411118, 0.034434083849191666, -0.03235827013850212, -0.008539835922420025, -0.015644919127225876, 0...
90076560a7c9cfab38e76aef193c708921824193
subsection
4
106
Setting and Background
We consider reinforcement learning (RL) in a common special case of contextual decision processes , , sometimes referred to as rich observation MDPs . We assume an H-step process where in each episode, a random trajectory s_1, x_1, a_1, r_1, s_2, x_2, \ldots , s_H, x_H, a_H, r_H is generated. For each time step (or lev...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 150, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016....
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0509817935526371, -0.025918032974004745, 0.0013147796271368861, 0.013401408679783344, -0.02248568646609783, 0.00258570141158998, 0.01085384376347065, 0.03935758024454117, 0.02698587439954281, 0.0042904336005449295, -0.03975420445203781, -0.03377429395914078, 0.004675619304180145, 0.0493...
713af3062c0e0a0e00d7dc216bf024008bcf2135
subsection
5
106
Setting and Background
Let \pi ^\star denote the optimal policy, which maximizes V^\pi , with optimal value function g^\star defined as g^\star (x) := \mathbf {E}[\sum _{h^{\prime }=h}^H r_{h^{\prime }} | x_h=x, a_{h:H} \sim \pi ^\star ]. As is standard, g^\star satisfies the Bellman equation: \forall x at level h,g^\star (x) = \max _{a\in ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1473, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0635870099067688, -0.013684633187949657, -0.04613414779305458, 0.003280040342360735, -0.02968817763030529, 0.015324653126299381, 0.018108872696757317, 0.02596571296453476, 0.061146050691604614, 0.01684262603521347, -0.022090690210461617, 0.014416920952498913, 0.03103070706129074, 0.0290...
595c4f5d7237eba9456c43bb0c1c6da66d54f8d0
subsection
6
106
Function Classes and Optimization Oracles
As can be rich, the agent must use function approximation to generalize across observations. To that end, we assume a given value function class \subset (\rightarrow [0,1]) and policy class \Pi \subset (\rightarrow ). Our algorithm is agnostic to the specific function classes used, but for the guarantees to hold, they ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.48550/arxiv.1610.09512", "end": 627, "openalex_id": "https://openalex.org/W2545659366", "raw": "Nan Jiang, Akshay Krishnamurthy, Alekh Agarwal, John Langford, and Robert E. Schapire. Contextual decision processes with low Bellman rank ar...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.05795902758836746, -0.008767103776335716, -0.051702260971069336, 0.031177012249827385, -0.003387810429558158, -0.03509893640875816, 0.02281430922448635, -0.004841363988816738, 0.03592300042510033, 0.040775809437036514, -0.010285514406859875, -0.02150191366672516, 0.02469133958220482, 0....
3879b65db9232bed1fad45d6a83ec51abfa677fd
subsection
7
106
Function Classes and Optimization Oracles
Formally, for a program of the form\textstyle \max _{g \in } o(g) \textrm {, ~ subject to } h_j(g) \le c_j, ~~ \forall j \in [m],with constants \lbrace c_j\rbrace _{j\in [m]}, an LP oracle with approximation parameters \epsilon _{\textrm {sub}},\epsilon _{\textrm {feas}} returns a function \hat{g} that is at most \epsi...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 844, "openalex_id": "https://openalex.org/W2115044435", "raw": "Zheng Wen and Benjamin Van Roy. Efficient exploration and value function generalization in deterministic systems. In Advances in Neural Information Processing Systems...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.041196372359991074, 0.02625887282192707, -0.07122395187616348, 0.01502904761582613, -0.008414740674197674, -0.013068404980003834, 0.016829481348395348, -0.01311417855322361, 0.0025289994664490223, 0.010558477602899075, -0.023130999878048897, -0.014411102049052715, 0.01293871272355318, 0...
02cd41f30e6251711aaca4dec3390ca5acc80bd0
subsection
8
106
Body
In this section we propose and analyze a new algorithm, Valor (Values stored Locally for RL) shown in Algorithm  (with & as subroutines). As we will show, this algorithm is oracle-efficient and enjoys a polynomial sample-complexity guarantee in the deterministic hidden-state dynamics setting described earlier, which ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 350, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016....
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.005793004296720028, -0.01909898966550827, -0.010510547086596489, 0.03163842484354973, -0.0013681635027751327, 0.013805580325424671, 0.0507679246366024, 0.02860272116959095, 0.01273011788725853, 0.0027859050314873457, -0.017100613564252853, 0.012203827500343323, 0.018961697816848755, 0.0...
92a5b8d0d5a403cf79afc6007553853fe54e8a27
subsection
9
106
Body
\in T (a_{1:h}) *Alg. failure Main Algorithm Valor [H] 2mm2mm myfunFunction \hat{V}^{\star } \leftarrow V of the only dataset in _1 h=1:H CSC-oracle \hat{\pi }_h \leftarrow \limits _{\pi \in \Pi _h}\hspace{-3.33328pt}\sum \limits _{(D,V, \lbrace V_{a}\rbrace ) \in _h} \hspace{-10.0pt}V_D(\pi ; \lbrace V_a\rbrace...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.058844391256570816, -0.005695972125977278, -0.05368635058403015, 0.011147776618599892, -0.009194435551762581, -0.008698470890522003, -0.006214828230440617, 0.008637428283691406, 0.010667071677744389, 0.022844931110739708, -0.04929133132100105, -0.01817522570490837, 0.05377791076898575, ...
91b799b0b6b39b365f0c16277164007e2bd7547b
subsection
10
106
Body
\forall (D, V,\_) \in _{h+1}: \vspace*{2.84526pt} ~~|V - \hat{\mathbf {E}}_{D}[g(x_{h+1})]| \le \phi _{h+1} |V_{opt} - V_{pes}| \le 2\phi _{h+1} + 4 \epsilon _{\textrm {stat}}+ 2\epsilon _{\textrm {feas}} V_a \leftarrow (V_{opt} + V_{pes})/2 *consensus among remaining functions V_a \leftarrow (p \circ a) *no conse...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1020, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.02606584131717682, 0.006085336208343506, -0.04306663200259209, 0.03912928327918053, -0.023501988500356674, 0.01614617183804512, 0.05042244866490364, -0.0006905616610310972, 0.028080297634005547, 0.03708430752158165, -0.019564641639590263, 0.008302001282572746, 0.036504387855529785, 0.03...
084050a7ced58075b0c050bc77ca1792ec83f816
subsection
11
106
Body
For convenience, we also define _{H+1} to be the singleton \lbrace x\mapsto 0\rbrace . This notation also allows our algorithms to handle more general non-stationary function classes.Details of depth-first search exploration.  Valor maintains many data sets collected at paths visited by . Each data set D is collected f...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 2526, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.04567313939332962, -0.034140445291996, -0.034079425036907196, 0.034140445291996, -0.012638676911592484, 0.00338849239051342, 0.06449767202138901, 0.0004249917110428214, 0.022623002529144287, 0.0379236564040184, -0.015262517146766186, -0.013111578300595284, 0.01104454044252634, 0.0067350...
b0afe450420c462128b8079bc35d742fd0d52cec
subsection
12
106
Body
Assuming P \ne NP, even with algorithm parameter \phi =0 and perfect evaluation of expectations, Olive is not oracle-efficient, that is, it cannot be implemented with polynomially many basic arithmetic operations and calls to CSC, LP, and LS oracles. The assumptions of perfect evaluation of expectations and \phi =0 ar...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0331183522939682, 0.010339945554733276, -0.047311931848526, 0.028600824996829033, -0.018131151795387268, -0.013483899645507336, 0.04151240363717079, 0.0075660934671759605, 0.02107670158147812, 0.05109688267111778, -0.03565182909369469, -0.0038936955388635397, 0.005311145912855864, 0.046...
222172858666da84dc0424e426065ae6c3d5be53
subsection
13
106
Body
At the same time, oracles are implementable in polynomial time: For tabular value functions = (\rightarrow [0,1]) and policies \Pi = (\rightarrow ), the CSC, LP, and LS oracles can be implemented in time polynomial in ||, K = || and the input size. Both proofs are in Appendix . Proposition REF implies that if Olive c...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1016/0041-5553(80)90061-0", "end": 1367, "openalex_id": "https://openalex.org/W2033040247", "raw": "Leonid G Khachiyan. Polynomial algorithms in linear programming. USSR Computational Mathematics and Mathematical Physics, 1980.", "...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0070229265838861465, -0.00046039081644266844, -0.05865536257624626, 0.02230856940150261, -0.005363516975194216, -0.028946207836270332, 0.030563654378056526, 0.004859972279518843, 0.03509556129574776, 0.04141275957226753, 0.0032711350359022617, 0.0007629469619132578, -0.005237631034106016,...
bef2834b85bbbac7d1dc950208fc5224c8de542b
subsection
14
106
Body
We construct a family of MDPs as shown in Figure REF that encodes the 3-SAT problem for this formula as follows: For each variable x_i there are two terminal states x_i^1 and x_i^0 corresponding to the Boolean assignment to the variable. For each variable, the reward in either x_i^1 or x_i^0 is 1 and 0 in the other. Th...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.04052574560046196, -0.01776053011417389, -0.05724871903657913, -0.01323647703975439, -0.025496430695056915, 0.0368637815117836, 0.028578585013747215, 0.05401398241519928, 0.00580192543566227, 0.017638463526964188, -0.013160185888409615, -0.016951845958828926, 0.013747626915574074, -0.00...
0b4ab865ef8f4a723e20baf469fde760aeb2603c
subsection
15
106
Body
We list these constraints in the following writing out the constraints for each optimal action that are implied by the indicator of the original constraints in Problem (REF ): From initial state:f(s_0, \texttt {[try c_j]}) &= \max _b f(c_1, b) - 1/m & \textrm {if } \pi _{f}(s_0) =& \texttt {[try c_j]}\\ f(s_0, \texttt...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.026253588497638702, 0.0008585838950239122, -0.04322683811187744, 0.0036518434062600136, -0.012065965682268143, 0.03648027405142784, 0.022941362112760544, 0.03669396787881851, 0.0000538403655809816, 0.022391868755221367, -0.04011303931474686, -0.005899425130337477, -0.02523091807961464, ...
50f80f5846c59678735459daa029972def3d2797
subsection
16
106
What is new compared to
The overall structure of Valor is similar to Lsvee  . The main differences are in the pruning mechanism, where we use a novel state-identity test, and the policy optimization step in Algorithm .Lsvee uses a Q-value function class \subset (\times \rightarrow [0,1]) and a state identity test based on Bellman errors on da...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 53, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016."...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.05168153718113899, 0.0016303113661706448, -0.03070976212620735, 0.02359706349670887, -0.03168661519885063, 0.03324346989393234, 0.05839738994836807, 0.03318241611123085, 0.012767751701176167, 0.06526587903499603, -0.03376242145895958, -0.007845336571335793, 0.0096311429515481, 0.0479267...
803889b8f0b2186431cea827b4dbe6c7f8417a56
subsection
17
106
Computational and Sample Complexity of
Valor requires two types of nontrivial computations over the function classes. We show that they can be reduced to CSC on \Pi and LP on (recall Section REF ), respectively, and hence Valor is oracle-efficient.First, Lines  in and  in involve optimizing V_D(\pi ; \lbrace V_a\rbrace ) (Eq. (REF )) over \Pi , which can be...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.04173814877867699, 0.014515296556055546, -0.04036518558859825, 0.047016438096761703, -0.011410868726670742, 0.01798584684729576, 0.03227994218468666, -0.0013729655183851719, 0.047352053225040436, 0.019999530166387558, -0.017512938007712364, -0.014568689279258251, 0.03206637129187584, 0....
a9f9ec35108ee826bcc5a8e44416a8facc28a4c0
subsection
18
106
Computational and Sample Complexity of
Then for any \epsilon , \delta \in (0, 1), with probability at least 1- \delta , Valor makes O\left( \frac{MH^2}{\epsilon }\log \frac{MH}{\delta }\right) CSC oracle calls and at most O\left( \frac{MKH^2}{\epsilon }\log \frac{MH}{\delta }\right) LP oracle calls with required accuracy \epsilon _{\textrm {feas}}= \epsilon...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.48550/arxiv.1610.09512", "end": 1058, "openalex_id": "https://openalex.org/W2545659366", "raw": "Nan Jiang, Akshay Krishnamurthy, Alekh Agarwal, John Langford, and Robert E. Schapire. Contextual decision processes with low Bellman rank a...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.05415866896510124, 0.010091819800436497, -0.03990959748625755, 0.014821168035268784, 0.019573399797081947, 0.008619618602097034, 0.011106341145932674, 0.0041458080522716045, 0.04592044651508331, 0.0050535378977656364, -0.053029727190732956, 0.007692819461226463, 0.013692226260900497, 0....
c84ff247083bf44f9cfcb3e216932c8299540335
subsection
19
106
Toward Oracle-Efficient PAC-RL with Stochastic Hidden State Dynamics
Valor demonstrates that provably sample- and oracle-efficient RL with rich stochastic observations is possible and, as such, makes progress toward reliable and practical RL in many applications. In this section, we discuss the natural next step of allowing stochastic hidden-state transitions.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0007616591756232083, 0.03766611963510513, -0.007573667913675308, 0.041939422488212585, -0.022770604118704796, 0.035315804183483124, 0.040962669998407364, 0.013186194933950901, 0.022205917164683342, -0.006890702527016401, -0.013552477583289146, 0.015811223536729813, 0.02860061079263687, ...
5b5ccd5bb3952a2475c38c86a90e934b11114f7b
subsection
20
106
Computational Barriers with Decoupled Learning Rules.
One factor contributing to the computational intractability of Olive is that (REF ) involves optimizing over policies and values jointly. It is therefore promising to look for algorithms that separate optimizations over policies and values, as in Valor. In Appendix , we provide a series of examples that illustrate some...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1016/b978-1-55860-377-6.50040-2", "end": 539, "openalex_id": "https://openalex.org/W1547105496", "raw": "Geoffrey J Gordon. Stable function approximation in dynamic programming. In International Conference on Machine Learning, 1995.", ...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.021114984527230263, -0.015180209651589394, -0.025798728689551353, 0.017117783427238464, -0.006098782643675804, 0.005686857271939516, 0.024639233946800232, 0.03460172563791275, -0.00812408234924078, 0.04711204767227173, -0.020825112238526344, -0.006049199029803276, 0.015653161332011223, ...
3222078c983cea35d1cc5073f54c04947b673aed
subsection
21
106
Alternative Algorithms.
An important element of Valor is that it explicitly stores value estimates of the hidden states, which we call “local values." Local values lead to statistical and computational efficiency under weak realizability conditions, but this approach is unlikely to generalize to the stochastic setting where the agent may not ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 791, "openalex_id": "https://openalex.org/W2165421048", "raw": "J Andrew Bagnell, Sham M Kakade, Jeff G Schneider, and Andrew Y Ng. Policy search by dynamic programming. In Advances in Neural Information Processing Systems, 2004."...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.03518076241016388, -0.001903209020383656, -0.012265972793102264, 0.006510576698929071, -0.00922236405313015, 0.014218764379620552, 0.051016680896282196, 0.028879957273602486, 0.033838216215372086, 0.011442138813436031, -0.025889744982123375, 0.013974664732813835, -0.017971785739064217, ...
39374ebd0a540aa1bf61b040631e53f7b86c71d0
subsection
22
106
Conclusion
This paper describes new RL algorithms for environments with rich stochastic observations and deterministic hidden state dynamics. Unlike other existing approaches, these algorithms are computationally efficient in an oracle model, and we emphasize that the oracle-based approach has led to practical algorithms for many...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 702, "openalex_id": "https://openalex.org/W2480004914", "raw": "Matthew Johnson, Katja Hofmann, Tim Hutton, and David Bignell. The Malmo Platform for artificial intelligence experimentation. In International Joint Conference on Ar...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.017624767497181892, 0.027970734983682632, -0.014687306247651577, 0.008774234913289547, -0.023392872884869576, -0.00416966900229454, 0.0019331928342580795, 0.02621588669717312, 0.024979865178465843, 0.018036775290966034, -0.01828092709183693, -0.003368543228134513, 0.004287930205464363, ...
f70e24fab61cc42a6d0e3fb7a546e714e4fa3380
subsection
23
106
Additional Notation and Definitions
In the next few sections we analyze the new algorithms for the deterministic setting. We will adopt the following conventions:In the deterministic setting (which we focus on here), a path p always deterministically leads to some state s, so we use them interchangeably, e.g., V^\star (p) \equiv V^\star (s), x \sim p \Le...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.003918685019016266, -0.013881378807127476, -0.05787596106529236, 0.04154492914676666, -0.02930428460240364, -0.013209821656346321, 0.01800992526113987, 0.039408158510923386, 0.03437148407101631, 0.05213720723986626, -0.04004918783903122, 0.011843814514577389, 0.018162550404667854, 0.003...
cfe43b323ff147982828442d155452c9ec6e8ce2
subsection
24
106
Least-Squares (LS) Oracle
The least-squares oracle takes as inputs a parameter \epsilon _{\textrm {sub}} and a sequence \lbrace (x^{(i)}, v^{(i)})\rbrace _{i\in [n]} of observations x^{(i)}\in and values v^{(i)}\in . It outputs a value function \hat{g} \in whose squared error is \epsilon _{\textrm {sub}} close to the least-squares fit\min _{g \...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.03448543697595596, -0.0004873356956522912, -0.06781119108200073, 0.028473375365138054, 0.0288395918905735, 0.007099269889295101, 0.00941483024507761, -0.04470899701118469, -0.032898496836423874, 0.0001771479583112523, -0.044251225888729095, -0.016708653420209885, -0.01904328726232052, 0...
7b380a2424809d8101dfe6850e6cceb9a5d3b2dd
subsection
25
106
Multi Data Set Classification Oracle
The multi data set classification oracle receives as inputs a parameter \epsilon _{\textrm {feas}}, m scalars that are upper bounds on the allowed cost \lbrace U_j\rbrace _{j \in [m]} \in ^{m}, and m cost-sensitive classification data sets D_1, \dots D_m, each of which consists of a sequence of observations \lbrace x_j...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1571, "openalex_id": "", "raw": "Sanjeev Arora, Elad Hazan, and Satyen Kale. The multiplicative weights update method: a meta-algorithm and applications. Theory of Computing, 2012.", "source_ref_id": "da3a97e7c85de061b9170cb...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.035072311758995056, -0.008111807517707348, -0.08913068473339081, 0.0490523986518383, -0.0008818634087219834, -0.024236219003796577, -0.024220958352088928, -0.0020184130407869816, 0.019672850146889687, 0.02240476757287979, -0.023793619126081467, 0.014315852895379066, 0.033851344138383865, ...
537d0975ee270bfc58990c8310de677ca7db6968
subsection
26
106
Assumptions on the Function Classes
While Valor only requires realizability of the policy and the value function classes, our other algorithms require stronger assumptions which we introduce below.[Policy realizability] \pi ^\star \in \Pi .[Value realizability] g^\star \in .[Policy-value completeness] At each level h, \forall g^{\prime } \in _{h+1}, t...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1551, "openalex_id": "https://openalex.org/W2117355432", "raw": "Rémi Munos and Csaba Szepesvári. Finite-time bounds for fitted value iteration. Journal of Machine Learning Research, 2008.", "source_ref_id": "6f51d77deaa045f...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.02473078854382038, -0.011450003832578659, -0.007666391786187887, 0.03020787239074707, -0.005171953700482845, -0.00970313511788845, 0.03606636822223663, 0.04122306779026985, 0.03101646713912487, 0.030818132683634758, -0.031275827437639236, -0.01155679952353239, 0.01858241856098175, 0.039...
65b8765db527d09b733181e97f11b244a5abea56
subsection
27
106
Analysis of
A state s \in _h is called learned if there is a data set in _h that is sampled from a path leading to that state. The set of all learned states at level h is ^{\textrm {learned}}_h and ^{\textrm {learned}}:= \bigcup _{h \in [H]} ^{\textrm {learned}}_h.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ 0.008425199426710606, -0.027397161349654198, -0.028373997658491135, 0.018483545631170273, 0.00465141236782074, -0.03330396115779877, 0.019780904054641724, 0.04240073263645172, -0.03171660378575325, -0.004288915079087019, -0.014393049292266369, -0.01642303355038166, -0.004853647667914629, 0...
e14532426725864514c0f577b2e34ab7f95fc4bf
subsection
28
106
Concentration Results
We now define an event that holds with high probability and will be the main concentration argument in the proof. This event uses a parameter \epsilon _{\textrm {stat}} whose value we will set later.[Deviation Bounds] Let denote the event that for all h \in [H] the total number of calls to (p) at level h is at most T_...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ 0.004272399470210075, -0.041015032678842545, -0.034453850239515305, 0.013770858757197857, -0.026092153042554855, -0.0335688516497612, -0.04156434163451195, -0.012664612382650375, 0.01425150316208601, 0.030639206990599632, -0.03997744992375374, 0.03573556989431381, 0.0011014780029654503, 0....
c1aff27ce2e86ad8510d94a201af90c64cda30c1
subsection
29
106
Concentration Results
The current path is denoted by p_j at level h_j and data sets D^{\prime }_{a}, \tilde{D} collected are denoted by D^{\prime }_{j,a} and \tilde{D}_{j} respectively. Consider a fix a \in and g \in and defineY_{i,j} = {\left\lbrace \begin{array}{ll} 0 & \textrm {if } j > N_{dfs}\\ g(x_{h+1}^{(i,j)}) - \mathbf {E}_{p_j \ci...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.05967530235648155, -0.0009243226959370077, -0.021981623023748398, -0.0000657251148368232, -0.007947553880512714, 0.014987470582127571, -0.01218828372657299, 0.02089856006205082, 0.022378237918019295, 0.052688777446746826, -0.0011869848240166903, 0.013012022711336613, 0.005300912074744701,...
a5879782468396b2d07437c1ccea127f5952ab6e
subsection
30
106
Concentration Results
With a union bound over and , the following statement holds: Given a fix j \in , with probability at least 1-\delta ^{\prime }, if j \le N_{dfs} then for all g \in _{h+1} and a \in\left| \hat{\mathbf {E}}_{D^{\prime }_{j,a}}[g(x_{h+1})] - \mathbf {E}_{p_j\circ a}[g(x_{h+1})]\right| \le \sqrt{\frac{\log (2K||/\delta ^{\...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.48550/arxiv.1610.09512", "end": 1366, "openalex_id": "https://openalex.org/W2545659366", "raw": "Nan Jiang, Akshay Krishnamurthy, Alekh Agarwal, John Langford, and Robert E. Schapire. Contextual decision processes with low Bellman rank a...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.026589324697852135, -0.006174428388476372, -0.027062226086854935, 0.00558330025523901, -0.02291288785636425, 0.007051586639136076, 0.01036571804434061, -0.004195102956146002, -0.01604817435145378, 0.03182176128029823, -0.02330951578915119, 0.003992975223809481, 0.00437053432688117, 0.01...
cd97ea1d91caea52915244e5832893801962b0ee
subsection
31
106
Concentration Results
As such, Bernstein's inequality with a union bound over \pi \in \Pi gives that with probability 1-\delta ^{\prime },\left| (\hat{\mathbf {E}}_{\tilde{D}_j}- \mathbf {E}_{p_j})[K \lbrace \pi (x_h)=a_h\rbrace (r_h + V_{a_h})]\right| \le \sqrt{\frac{4 K\log (2 |\Pi |/\delta ^{\prime })}{n_{\textrm {train}}}} + \frac{4K}{3...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.011862057261168957, 0.03188642859458923, -0.042169421911239624, 0.010603382252156734, -0.04152864217758179, 0.03316798806190491, 0.01027536392211914, 0.04360354691743851, -0.014074273407459259, 0.0504385344684124, -0.05971458554267883, 0.023266414180397987, 0.023281671106815338, 0.02749...
1590afcf94a5878e760bf2d07fab11ec043c761e
subsection
32
106
Concentration Results
This means that D^{\prime }_a and D are two data sets sampled from the same distribution, and as such, we haveV_{opt} - V_{pes} & = \hat{\mathbf {E}}_{D_a^{\prime }}[g_{opt}(x_{h+1}) - g_{pes}(x_{h+1})] \le \mathbf {E}_s[g_{opt}(x_{h+1}) - g_{pes}(x_{h+1})] + 2\epsilon _{\textrm {stat}}\\ & \le \hat{\mathbf {E}}_D[g_{o...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.008569308556616306, -0.006161813624203205, -0.029698796570301056, -0.018252704292535782, -0.008065680973231792, -0.004982865881174803, -0.00550938630476594, 0.00030403686105273664, -0.011942090466618538, 0.03656645491719246, -0.0748421922326088, -0.01701652631163597, 0.02785216085612774, ...
562d1e50ff8eb18a6b6bc1df1d613690f4ecba02
subsection
33
106
Bound on Oracle Calls
[Proof of Theorem REF ] Consider event from Definition REF which by Lemma REF has probability at least 1 - \delta /2. Valor requires two types of nontrivial computations over the function classes. We show that they can be reduced to CSC on \Pi and LP on (recall Sec. REF ), respectively, and hence Valor is oracle-effici...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.024996044114232063, 0.01658772863447666, -0.057988379150629044, 0.04065290465950966, -0.0038550826720893383, 0.023058010265231133, 0.0013247673632577062, -0.011216172948479652, 0.02231026627123356, 0.028017543256282806, -0.018815703690052032, 0.005901843775063753, 0.03644111752510071, 0...
e02a5f2052678b788a25510ec663b0c9e6c871f4
subsection
34
106
Bound on Oracle Calls
In addition, there at at most MH calls to the CSC oracle in . The statement follows with realizing that T_{\max }= MH n_{\textrm {exp}}+ M = O\left(\frac{MH}{\epsilon } \ln \left( \frac{MH}{\delta }\right) \right).
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0368327796459198, 0.003080206224694848, -0.057339515537023544, 0.029722558334469795, 0.006179484538733959, 0.011145959608256817, -0.0598723404109478, -0.0161276925355196, 0.008620762266218662, -0.009391291067004204, -0.011725762858986855, 0.009559128433465958, -0.017638232558965683, 0.0...
3bb4f376d69930acce25e6c105b87464f8c38577
subsection
35
106
Depth First Search and Estimated Values
In this section, we show that in the high-probability event (Definition REF ), produces good estimates of optimal values on learned states. The next lemma first quantifies the error in the value estimate at level h in terms of the estimation error of the values of the next time step \lbrace V_a\rbrace _a. [Error propa...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.019471550360322, -0.03598269075155258, -0.02337806671857834, 0.01039194781333208, -0.00434142304584384, 0.05084577202796936, 0.01059795543551445, 0.02230987884104252, 0.018174463883042336, 0.047885362058877945, -0.03561645373702049, 0.014733370393514633, 0.007389575242996216, 0.04928926...
b6704986cdf8f91a2a24cfde4c9d76d85320a726
subsection
36
106
Depth First Search and Estimated Values
The third inequality uses that \pi ^\star is the global and point-wise maximizer of the long-term expected reward, which is precisely r_h+g^\star .Similarly, we can lower bound \tilde{V} by\tilde{V} & = \hat{\mathbf {E}}_{\tilde{D}} [ K \lbrace \tilde{\pi }(x_h) = a_h\rbrace (r_h + V_{a_h})] - \epsilon _{\textrm {sub}}...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.03305116295814514, -0.0013437524903565645, -0.033905670046806335, 0.019928354769945145, -0.009216482751071453, 0.02188151702284813, 0.01143667846918106, -0.006229518447071314, 0.04504479467868805, 0.025101182982325554, -0.05880848318338394, 0.015808403491973877, 0.020965972915291786, 0....
81bc99c1945e2b33598fd96c0295359f4c260812
subsection
37
106
Depth First Search and Estimated Values
Since g^\star _{h+1} is feasible for both V_{opt} and V_{pes}, we haveV_{pes} - \epsilon _{\textrm {sub}}=& \hat{\mathbf {E}}_{D_a^{\prime }}[g_{pes}(x_{h+1})] - \epsilon _{\textrm {sub}}\le \hat{\mathbf {E}}_{D_a^{\prime }}[g^\star (x_{h+1})] \le \hat{\mathbf {E}}_{D_a^{\prime }}[g_{opt}(x_{h+1})] + \epsilon _{\textrm...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.02919928729534149, 0.010707423090934753, -0.03828113526105881, 0.011058486066758633, -0.01253905612975359, -0.014317266643047333, 0.024665994569659233, 0.02498653158545494, 0.032542020082473755, -0.0038426141254603863, -0.043806564062833786, -0.006990734022110701, 0.03702951967716217, 0...
308e20420d9df5b1dd506dbd72bc10428f2dcb17
subsection
38
106
Depth First Search and Estimated Values
Therefore,\hat{\mathbf {E}}_{D^{\prime }}[g^\star (x_{h+1})] - V_a & \le V_{opt} - V_a + \epsilon _{\textrm {sub}}= \frac{V_{opt} - V_{pes}}{2} + \epsilon _{\textrm {sub}}\le \phi _{h+1} + 2\epsilon _{\textrm {stat}}+ \epsilon _{\textrm {feas}}+ \epsilon _{\textrm {sub}}.\\ V_a - \hat{\mathbf {E}}_{D^{\prime }}[g^\star...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.008942443877458572, -0.0037120296619832516, -0.02102847583591938, -0.03247358277440071, -0.004074458032846451, -0.002132604829967022, 0.015992630273103714, 0.019761884585022926, 0.031206991523504257, 0.012894820421934128, -0.036441221833229065, 0.01666407473385334, 0.028246523812413216, ...
65ac6a6383fb834012574f6798f9e895b91bf908
subsection
39
106
Depth First Search and Estimated Values
For h=H+1 the statement holds trivially since _{H+1} = \lbrace g^\star _{h+1} \rbrace the constant 0 function is the only function in _{H+1} and therefore the algorithm always returns on Line  and never calls level H+1 recursively.Consider now some data set (\tilde{D},\tilde{V}, \lbrace V_a\rbrace ) \in _h at level h ...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.04447284713387489, 0.014987073838710785, -0.027822237461805344, -0.0006095176213420928, 0.013468526303768158, 0.01455974392592907, 0.010774821043014526, 0.04184781759977341, 0.03641462326049805, 0.020374484360218048, -0.04294666647911072, -0.004143575206398964, 0.02454095333814621, 0.02...
3b69289c8eeec3aa48e9026e875098e4b7beed62
subsection
40
106
Policy Performance
In this section, we bound the quality of the policy returned by in the good event by using the fact that produces accurate estimates of the optimal values (previous section). Before we state the main result of this section in Proposition REF , we prove the following helpful lemma. This Lemma is essentially Lemma 4.3 in...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 323, "openalex_id": "https://openalex.org/W112666333", "raw": "Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv:1406.5979, 2014.", "source_ref_id": "97e1333f7...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.05104175955057144, 0.00714294845238328, -0.015956271439790726, 0.013370622880756855, -0.027992120012640953, -0.026527682319283485, -0.024514079093933105, 0.045489102602005005, 0.05095023289322853, 0.0646793395280838, -0.05095023289322853, 0.04564164578914642, 0.012600267305970192, 0.028...
f9070632f3f794cacc932573543c4b436325289e
subsection
41
106
Policy Performance
Then the policy \hat{\pi } = \hat{\pi }_{1:H} returned by satisfiesV^{\hat{\pi }} \ge V^\star - p_{ul}^{\hat{\pi }} - 2 H^2 T_{\max }(7\epsilon _{\textrm {stat}}+ 3\epsilon _{\textrm {sub}}+ 2\epsilon _{\textrm {feas}})where p_{ul}^{\hat{\pi }}= \mathbf {P}( \exists h \in [H] ~:~ s_h \notin ^{\textrm {learned}}~|~ a_{1...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.012783506885170937, -0.012882662937045097, -0.014980195090174675, 0.02388898655772209, -0.013660656288266182, -0.02745860628783703, -0.0036592406686395407, 0.04442192241549492, 0.009457964450120926, 0.01659720204770565, -0.040242113173007965, 0.01668873056769371, 0.008275718428194523, 0...
187940d900b822a0ee61f0aabf52267027b02b1e
subsection
42
106
Policy Performance
For a state s \in ^{\textrm {learned}} at level h, we have& V^\star (s) - Q^\star (s, \hat{\pi }_h)\\ & = \mathbf {E}_s[K(\lbrace \pi ^\star _h(x_h) = a_h\rbrace - \lbrace \hat{\pi }_h(x_h) = a_h\rbrace )(r_h + g^\star _{h+1}(x_{h+1})]\\ & \le \sum _{s \in ^{\textrm {learned}}_h} \mathbf {E}_s[K(\lbrace \pi ^\star _h(x...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.024310637265443802, -0.012918364256620407, -0.027042340487241745, -0.006688096094876528, -0.020098624750971794, 0.022311458364129066, 0.019900232553482056, 0.052436504513025284, 0.015085414052009583, 0.02815638855099678, -0.015466936863958836, -0.001472678268328309, -0.011689860373735428,...
689c7e526079649670b062e592f35a3400698e6f
subsection
43
106
Policy Performance
The third inequality uses the deviation bound that holds in event .Since per call, only one data set can be added to _{h}, the magnitude |_{h}| \le T_{\max } is bounded by the total number of calls to at each level. Using Lemma REF , the suboptimality of \hat{\pi } is therefore at mostV^\star - V^{\hat{\pi }} & \le p_{...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1351, "openalex_id": "https://openalex.org/W2963971282", "raw": "Akshay Krishnamurthy, Alekh Agarwal, and John Langford. PAC reinforcement learning with rich observations. In Advances in Neural Information Processing Systems, 2016...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.022840403020381927, -0.0024945931509137154, -0.043483734130859375, -0.0002944211009889841, -0.00016795114788692445, 0.01629495620727539, -0.011076908558607101, 0.004550534766167402, 0.010542898438870907, 0.02374059334397316, -0.08409906923770905, 0.04998340830206871, 0.012595024891197681,...
ab3be4ed087ad09aa7703929fdac50ce4b09c36c
subsection
44
106
Meta-Algorithm Analysis
Now that we have the main guarantees for and , we may turn to the analysis of . Consider running with and (Algorithm  + + ) with parametersn_{\textrm {exp}}\ge \frac{8}{\epsilon }\ln \left(\frac{4MH}{\delta }\right), \quad n_{eval} \ge \frac{32}{\epsilon ^2}\ln \left( \frac{8MH}{\delta }\right), \quad \epsilon _{\tex...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.028320586308836937, 0.0061531662940979, -0.038422003388404846, 0.05319265276193619, -0.01992817036807537, -0.0003328355378471315, 0.0036697741597890854, 0.014450212940573692, 0.010528665967285633, 0.011665456928312778, -0.03056364879012108, 0.014701984822750092, 0.0009570214315317571, 0...
633e42973816762d1ccfbef0161aefdc71c2be38
subsection
45
106
Meta-Algorithm Analysis
With these two bounds, if terminates, the termination condition implies thatV^\star - V^{\hat{\pi }^{(k)}} \le \hat{V}^{(k)} - \hat{V}^{\hat{\pi }^{(k)}} + \frac{\epsilon }{4} \le \frac{3}{4}\epsilon \le \epsilonand hence the returned policy is \epsilon -optimal.On the other hand, if the algorithm does not terminate in...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.01333326380699873, 0.01848960481584072, -0.01916084624826908, -0.013630745001137257, -0.014065524563193321, 0.014904574491083622, 0.015789391472935677, 0.007189130876213312, 0.02300521917641163, 0.04948867857456207, -0.02556813508272171, 0.015522420406341553, 0.002753608860075474, 0.034...
b271830553824f55ae8a6b5a495427e8373a2b8e
subsection
46
106
Meta-Algorithm Analysis
With this fact, since there are at most MH states, the algorithm must terminate and return a near-optimal policy after at most MH iterations.In a non-terminal iteration k, the probability that we do not hit an unlearned state in Line  is(1 - p_{ul}^{\hat{\pi }^{(k)}})^{n_{\textrm {exp}}} \le (1 - \epsilon /8)^{n_{\text...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.04409252479672432, -0.020367998629808426, -0.028042234480381012, -0.013174357824027538, -0.02425851672887802, 0.006003601476550102, 0.012762420810759068, 0.02163432538509369, -0.013891433365643024, 0.0308037381619215, -0.027081048116087914, 0.0005392370512709022, 0.024350058287382126, 0...
05a0ce94544f4711cea2f0782b4712789debd3c9
subsection
47
106
Proof of Sample Complexity: Theorem
We now have all parts to complete the proof of Theorem REF . For the calculation, we instantiate all the parameters as\epsilon _{\textrm {stat}}&= \epsilon _{\textrm {sub}}= \epsilon _{\textrm {feas}}= \frac{\epsilon }{2^6 7 H^2 T_{\max }},\\ \phi _h &= (H-h+1)(6\epsilon _{\textrm {stat}}+ 2\epsilon _{\textrm {sub}}+ \...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ 0.0035389899276196957, 0.003729770891368389, -0.01659032702445984, -0.008531731553375721, -0.009142231196165085, -0.009989299811422825, -0.012347354553639889, -0.01787237636744976, 0.020833298563957214, 0.011462129652500153, -0.0359584242105484, 0.0039415378123521805, 0.007429016754031181, ...
8743027387f24cce75a3815141b15038f7d1e5ee
subsection
48
106
Extension:
We note that Theorem REF suffers relatively high sample complexity compared to the original Lsvee. The issue is that Valor pools all the data sets together for policy optimization (Algorithm ). This implicitly weights all data sets uniformly, and allows some undesired trade-off: the policy that maximizes the objective ...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.03247160464525223, -0.030289538204669952, -0.03213590383529663, 0.02706984430551529, 0.004593023099005222, 0.019562311470508575, 0.007118423003703356, 0.02887043170630932, 0.04800548404455185, 0.03732403367757797, -0.04428223520517349, -0.0005021234974265099, 0.003717525629326701, 0.046...
9b95b72665a56439f6ce4bf956c5eb65d9365e78
subsection
49
106
Extension:
Using Proposition REF , we can bound the deviation of the optimal policy for each constraint as& V - \hat{\mathbf {E}}_{D}[K \lbrace \pi ^\star (x_h) = a_h\rbrace (r_h + V_{a_h})] \\ & \le V^\star (s) + \phi _h - 2\epsilon _{\textrm {stat}}- \hat{\mathbf {E}}_{D}[K \lbrace \pi ^\star (x_h) = a_h\rbrace (r_h + V_{a_h})]...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.01147464755922556, -0.027481170371174812, -0.032043565064668655, 0.015266469679772854, -0.033722035586833954, -0.006877921987324953, 0.00784304365515709, 0.031234845519065857, 0.0742800310254097, 0.04501357674598694, -0.04306044802069664, 0.016784723848104477, -0.006019612308591604, 0.0...
749664fd29fa850277c21154626178f07e7a7f04
subsection
50
106
Extension:
Then the policy \hat{\pi } = \hat{\pi }_{1:H} returned by in Algorithm REF satisfiesV^{\hat{\pi }} \ge V^\star - p_{ul}^{\hat{\pi }} - 32H^2(\epsilon _{\textrm {stat}}+\epsilon _{\textrm {feas}}+ \epsilon _{\textrm {sub}})where p_{ul}^{\hat{\pi }}= \mathbf {P}( \exists h \in [H] ~:~ s_h \notin ^{\textrm {learned}}~|~ a...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.01832328550517559, -0.040918443351984024, -0.016172092407941818, 0.02410557121038437, -0.012388432398438454, -0.024563271552324295, 0.03280188515782356, 0.03258829191327095, 0.005473337601870298, 0.026653436943888664, -0.06352885812520981, 0.008681057021021843, 0.02132885344326496, 0.01...
6c828cedf43e22fdfd2f4f3d46169601ddc7f372
subsection
51
106
Extension:
For a state s \in ^{\textrm {learned}}_h, we have& V^\star (s) - Q^\star (s, \hat{\pi }_h)\\ & = \mathbf {E}_s[K(\lbrace \pi ^\star _h(x_h) = a_h\rbrace - \lbrace \hat{\pi }_h(x_h) = a_h\rbrace ) (r_h + g^\star _{h+1}(x_{h+1})]\\ & \le \mathbf {E}_s[K(\lbrace \pi ^\star _h(x_h) = a_h\rbrace - \lbrace \hat{\pi }_h(x_h) ...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.013717329129576683, 0.0022468038368970156, -0.03405681625008583, 0.0012845691526308656, -0.0207972414791584, -0.00825481116771698, -0.0018977673025801778, 0.039885539561510086, 0.0517565980553627, 0.04351704567670822, -0.014121677726507187, 0.02157542109489441, 0.02387944422662258, 0.02...
3467297fc178473da0884130d46a7cfd333b5649
subsection
52
106
Extension:
Using Lemma REF , the suboptimality of \hat{\pi } is therefore at mostV^\star - V^{\hat{\pi }} & \le p_{ul}^{\hat{\pi }} + (1 -p_{ul}^{\hat{\pi }}) \sum _{h=1}^H (4 \phi _{h+1} + 16 \epsilon _{\textrm {stat}}+ 5 \epsilon _{\textrm {feas}}+ 6 \epsilon _{\textrm {sub}})\\ &\le p_{ul}^{\hat{\pi }} + 16H\epsilon _{\textrm ...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.04542330652475357, -0.0024688239209353924, -0.029320621863007545, 0.019201114773750305, -0.012950832024216652, -0.013767413794994354, 0.01697268709540367, 0.02871009334921837, 0.033914849162101746, 0.015507418662309647, -0.06728021800518036, 0.03119799681007862, -0.005986993201076984, 0...
95764aeee4163b98413f1e83c2c02def4bda1d1a
subsection
53
106
Extension:
Finally, we are ready to assemble all statements to the following sample-complexity bound: Consider a Markovian CDP with deterministic dynamics over M hidden states, as described in Section . When \pi ^\star \in \Pi and g^\star \in (Assumptions REF and REF hold), for any \epsilon , \delta \in (0, 1), the local value a...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.039305929094552994, -0.03353821113705635, -0.023238714784383774, -0.008376923389732838, -0.015029635280370712, 0.002513839863240719, 0.016647037118673325, 0.00005310677443048917, 0.0538015142083168, 0.017913494259119034, -0.04391399770975113, 0.032744769006967545, -0.003396926447749138, ...
a56154fead4b3f8bdddbbb8192425fac7da4f913
subsection
54
106
Extension:
For the sample complexity, since T_{\max } is an upper bound on the number of data sets we collect (because T_{\max } is an upper bound on the number of execution of at any level), and we also n_{\textrm {eval}} trajectories for each of the MH iterations of , the total sample complexity is& HT_{\max }n_{\textrm {train}...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.030501333996653557, -0.02777009829878807, -0.015182004310190678, 0.0019530619028955698, 0.02029353380203247, -0.011649709194898605, -0.017440233379602432, 0.005660827737301588, -0.0005869675660505891, -0.0060728020034730434, -0.04046500101685524, -0.002143790712580085, 0.01196250412613153...
621d541f73387545fafeaeb101a308791993386e
subsection
55
106
Alternative Algorithms
[Informal statement] Under Assumption REF or Assumptions REF +REF , there exist oracle-efficient algorithms with polynomial sample complexity in CDPs (contextual decision processes) with deterministic dynamics over small hidden states. These algorithms do not store or use local values.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.02498016133904457, -0.01863211765885353, -0.013817676343023777, 0.02900872752070427, -0.029939569532871246, -0.008217358961701393, 0.06268692761659622, 0.02342366985976696, 0.009155831299722195, -0.012665566988289356, -0.00894219521433115, -0.007282701320946217, -0.02655191160738468, 0....
a736dd5cda9e8a101213f66edee4a870cf75727c
subsection
56
106
Algorithm with Two-Sample State-Identity Test
See Algorithm  + REF . The algorithm uses a novel state identity test which compares two distributions using a two-sample test  in Line REF (recall that _h = for h \in [H] and _{H+1} = \lbrace x \mapsto 0 \rbrace ). Such an identity test mechanism is very different from the one used in the Valor algorithm, and the two ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.5555/2188385.2188410", "end": 218, "openalex_id": "https://openalex.org/W2212660284", "raw": "Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander J. Smola. A kernel two-sample test. Journal of Machine ...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.042140569537878036, -0.03371855989098549, -0.0022542611695826054, 0.010832658968865871, -0.007434103172272444, 0.019468272104859352, 0.008933129720389843, 0.04809090495109558, 0.02161954715847969, 0.02352670580148697, -0.041194621473550797, 0.003837202675640583, -0.012465187348425388, 0...
77c436755ea1d06f35258b7bea510f4d43c791d3
subsection
57
106
Algorithm with Two-Sample State-Identity Test
Given the novelty of the mechanism, we believe analyzing the two-sample test algorithm and understanding its computational and statistical properties enriches our toolkit for dealing with the challenges addressed in this paper.myfunFunction \hat{g}_{H+1} \leftarrow 0 h=H:1 \hat{\pi }_h \leftarrow _{\pi \in \Pi _h} \s...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.05619676411151886, -0.02547464706003666, -0.01130342110991478, -0.0029517030343413353, -0.012089015915989876, 0.0029231011867523193, 0.027793297544121742, 0.04869166016578674, -0.010998335666954517, 0.04057638347148895, -0.0565323606133461, -0.01777123473584652, 0.0012460839934647083, -...
dfe9c7dde5a0c4bda05243bef27245ea738416ae
subsection
58
106
Computational considerations
The two-sample test algorithm requires three types nontrivial computation. Line REF requires importance weighted policy optimization, which is simply a call to the CSC oracles. Line REF performs squared-loss regression on _h, which is a call to a LS oracle.The slightly unusual computation occurs on Line REF : we comput...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1198/jasa.2003.s269", "end": 922, "openalex_id": "https://openalex.org/W3144619878", "raw": "Bernhard Schölkopf and Alexander J. Smola. Learning with kernels: support vector machines, regularization, optimization, and beyond. MIT Press, ...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.07721023261547089, -0.020492754876613617, -0.040863439440727234, -0.009254546836018562, 0.002048512687906623, 0.0485234260559082, 0.02969389595091343, -0.011146655306220055, 0.013404978439211845, 0.028320591896772385, -0.06475893408060074, -0.006275238934904337, 0.007682875730097294, 0....
182d22c4681ed02fd06151bd5f4bcd61d4a6ed21
subsection
59
106
Sample complexity
Consider the same Markovian CDP setting as in Theorem REF but we explicitly require here that the process is an MDP over . Under Assumption REF , for any \epsilon , \delta \in (0, 1), the two-sample state-identity test algorithm (Algorithm +REF ) returns a policy \pi such that V^\star - V^{\pi } \le \epsilon with proba...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.03420931100845337, -0.04839961603283882, -0.037505123764276505, 0.003209983929991722, -0.02583245560526848, 0.025496771559119225, 0.032164689153432846, 0.03362949192523956, 0.01213042065501213, 0.02052253484725952, -0.054258838295936584, -0.01106233336031437, 0.010619839653372765, 0.027...
96cdb82c500e9de59cd7caa691c8fa617069dbdf
subsection
60
106
Sample complexity
\\ L_{^{\textrm {val}}_h}(g; \pi , g^{\prime }) :=& \sum _{D \in ^{\textrm {val}}_{h}} L_D(g; \pi , g^{\prime }).Also define V_s, V_{^{\textrm {learned}}_h}, L_s, L_{^{\textrm {val}}_h} as the population version of V_D, V_{^{\textrm {learned}}_h}, L_D, L_{^{\textrm {val}}_h}, respectively.Consider a Markovian contextua...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.019408274441957474, -0.03234712406992912, -0.038206227123737335, -0.007579450029879808, -0.011275719851255417, 0.011870784685015678, -0.0009278818615712225, 0.006786030251532793, 0.020094888284802437, 0.041318874806165695, -0.0676238015294075, 0.007972345687448978, 0.053708434104919434, ...
4f3718f344609d03897a166cfff4f26ac449ae42
subsection
61
106
Sample complexity
We say the deviation bound holds for a data set of n_{\textrm {test}} observations sampled in Line REF during a call to if for all g \in _h:|\hat{\mathbf {E}}_{D} [g(x_h)] - \mathbf {E}_{p}[g(x_h)]| & \le \tau _{val}, \qquad |\hat{\mathbf {E}}_{D} [\bar{r}] - V^{\hat{\pi }_{h+1:H}}(p)| \le \tau _{val}.We say the deviat...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.01773301512002945, -0.027194693684577942, -0.03339056670665741, -0.015794897451996803, -0.0069283898919820786, -0.04587387666106224, -0.013307392597198486, 0.016863152384757996, 0.009766893461346626, 0.0075426362454891205, -0.03470299392938614, 0.0013000268954783678, -0.01249094121158123,...
94a377ee947a6143636e05d3ecb85d1afac578b7
subsection
62
106
Concentration Results.
For our analysis we rely on the following concentration bounds that define the good event . This definition involves parameters \tau , \tau _L, \tau _V whose values we will set later. Let denote the event that for all h \in [H] the total number of calls to (p) at level h is at most T_{\max }= M(K+1)(1 + Hn_{\textrm {...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.0085448632016778, -0.047668129205703735, -0.050811417400836945, -0.009963920339941978, -0.019851546734571457, 0.013847255147993565, -0.01081077754497528, 0.013290313072502613, 0.004157991148531437, 0.05245935544371605, -0.03884860873222351, 0.016158945858478546, 0.008575379848480225, 0....
06d011be116855f01dcb8993a2ca5ac3cd2a4caf
subsection
63
106
Concentration Results.
By Hoeffding's inequality and a union bound, with probability 1-\delta ^{\prime }, for all g \in _{h}\left| \hat{\mathbf {E}}_{\tilde{D}}[g(x_h)] - \mathbf {E}_{s}[g(x_{h})]\right| \le \sqrt{\frac{\log (2||/\delta ^{\prime })}{2n_{\textrm {train}}}}.With \delta ^{\prime } = \frac{\delta }{6HT_{\max }} the choice for n_...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.48550/arxiv.1610.09512", "end": 600, "openalex_id": "https://openalex.org/W2545659366", "raw": "Nan Jiang, Akshay Krishnamurthy, Alekh Agarwal, John Langford, and Robert E. Schapire. Contextual decision processes with low Bellman rank ar...
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.01362299732863903, 0.014653307385742664, -0.02506326138973236, 0.008807248435914516, -0.012249249033629894, 0.007990632206201553, 0.006746626924723387, 0.008471443317830563, 0.0006935517303645611, 0.029337143525481224, -0.0422808974981308, 0.02207154408097267, 0.02718493901193142, 0.004...
ed89a19ab76164547a2f591939e40af0593c479a
subsection
64
106
Concentration Results.
Using a union bound, the deviation bounds (REF )–() hold for a single call to with probability 1 - 3 \delta ^{\prime }.Consider now the event ^{\prime } that these bounds hold for the first T_{\max } calls at each level h. Applying a union bound let us bound \mathbf {P}(^{\prime }) \ge 1 - 3 H T_{\max }\delta ^{\prime ...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.012956587597727776, -0.012605585157871246, -0.05164321884512901, -0.005078097339719534, -0.03983120620250702, 0.04617978259921074, -0.0037885429337620735, 0.008424071595072746, -0.012681889347732067, 0.030232038348913193, -0.059456851333379745, 0.016268223524093628, -0.0050933584570884705...
6abbfc13e07fb013d6f73317b62e6f55a4a5d1f4
subsection
65
106
Concentration Results.
The bound |^{\textrm {val}}_h| \le M follows from the fact that in no state can be added twice to ^{\textrm {val}}_h since as soon as it is in ^{\textrm {val}}_h once, d_{MMD} \le 2 \tau holds (see Eq.(REF )) and the current data set is not added to ^{\textrm {val}}_h.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.010190718807280064, 0.021586626768112183, -0.01766594685614109, -0.0036594290286302567, -0.03200617805123329, 0.016140390187501907, -0.011937480419874191, 0.0756676122546196, -0.0028280005790293217, -0.012784164398908615, -0.03289100155234337, -0.03255537897348404, 0.005438609514385462, ...
202147a7ae8de8b55dc8340bdb9e6f675ba5d3f1
subsection
66
106
Depth-first search and learning optimal values.
We now prove that and produce good value function estimates. In event , consider an execution of and let \lbrace \hat{g}_h,\hat{\pi }_h\rbrace _{h \in [H]} denote the learned value functions and policies. Then every state s in ^{\textrm {check}}_h satisfies\left|\mathbf {E}_{s} [\hat{g}_h(x_h)] - \mathbf {E}_{s}[g^\s...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.004051321651786566, -0.009285232052206993, -0.010628042742609978, 0.00014782363723497838, 0.006912424229085445, -0.03537086024880409, -0.0038395998999476433, 0.004844800569117069, -0.001926475903019309, 0.03573708236217499, 0.00867486372590065, -0.005676427856087685, 0.015045586042106152,...
862d95850555d0eea5af5f877815fbcfc4d4e1de
subsection
67
106
Base case:
Both statement holds trivially for h=H+1 since the LHS is 0 and the RHS is non-negative. In particular there are no actions, so Eq. (REF ) is trivial.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.03155405819416046, 0.0274801068007946, -0.023589253425598145, -0.005462453234940767, -0.01128347497433424, -0.01476235594600439, 0.04443202167749405, 0.032561104744672775, 0.023589253425598145, -0.0069386884570121765, -0.027144424617290497, -0.02868550829589367, 0.003379702102392912, 0....
ad029db95604b5539877a0f88710965653b055a1
subsection
68
106
Inductive case:
Assume that Eq. (REF ) holds on level h+1. For any learned s \in ^{\textrm {learned}}_h, we first show that \hat{\pi }_h achieves high value compared to \pi _{\hat{g}_{h+1}}^\star (recall its definition from Assumption REF ) under V_s(\cdot ; \hat{g}_{h+1}):V_s(\pi _{\hat{g}_{h+1}}^\star ; \hat{g}_{h+1}) - V_s(\hat{\pi...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.004497884307056665, -0.006397753953933716, -0.01593143679201603, -0.0019399271113798022, -0.0077902283519506454, -0.013680587522685528, 0.00935437809675932, 0.02818521484732628, 0.0398285910487175, -0.009239927865564823, -0.010453097522258759, 0.0020677295979112387, -0.030855713412165642,...
94be810965f84944fe5d20e1868872e0f8338934
subsection
69
106
Inductive case:
First we introduce and recall the definitions:&~ g_{\hat{\pi }_h,\hat{g}_{h+1}}(x) = \mathbf {E}[r + \hat{g}_{h+1}(x_{h+1}) \mid x_h = x, a_h = \hat{\pi }_h(x)], \\ &~ g_{\star ,\hat{g}_{h+1}}(x) = \mathbf {E}[r + \hat{g}_{h+1}(x_{h+1}) \mid x_h = x, a_h = \pi ^\star _{\hat{g}_{h+1}}(x)].Note that g_{\hat{\pi }_h,\hat{...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.02099880576133728, -0.023104790598154068, -0.040227364748716354, 0.0038113747723400593, 0.014558765105903149, -0.04437829181551933, 0.024493519216775894, 0.025042908266186714, 0.012658800929784775, 0.019976334646344185, -0.05200866982340813, 0.007233600132167339, -0.016435839235782623, ...
a2aacd566bac3b879ee17a990c1773168bf967be
subsection
70
106
Inductive case:
On the other hand, Assumption REF guarantees that g_{\star ,\hat{g}_{h+1}} \in _h, for any \hat{g}_{h+1}.The LHS of Eq.(REF ) can be bounded as\left|\mathbf {E}_{s} [g^\star (x_h)] - \mathbf {E}_{s}[\hat{g}_h(x_h)] \right| \le \left|\mathbf {E}_{s} [g^\star (x_h)] - \mathbf {E}_{s}[g_{\hat{\pi }_h, \hat{g}_{h+1}}(x_h)]...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.02178892306983471, 0.00043629342690110207, -0.007205755449831486, -0.010375677607953548, -0.011436132714152336, -0.00929233431816101, -0.015197315253317356, 0.02197202295064926, 0.06463436782360077, -0.0010633162455633283, -0.038054320961236954, 0.01747080869972706, -0.01222956646233797, ...
d8e26071c73619dda9ed4db09d428f6185e34a77
subsection
71
106
Inductive case:
The last inequality follows from the fact that if s \in ^{\textrm {learned}}_h \Rightarrow s \circ a \in ^{\textrm {check}}_{h+1} and we can therefore apply the induction hypothesis. We can use the same argument to lower bound the above quantity.
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.023724131286144257, 0.019482774659991264, 0.010427937842905521, 0.00743000116199255, -0.013708123005926609, -0.03219158574938774, -0.012891891412436962, 0.02482261136174202, 0.0030284500680863857, -0.0037550493143498898, -0.044152818620204926, -0.013746264390647411, -0.011488276533782482,...
a27e23501e9a5480c14edbc108637719b6efbfc2
subsection
72
106
Inductive case:
This gives\left|\mathbf {E}_{s} [g^\star (x_h)] - \mathbf {E}_{s}[g_{\hat{\pi }_h, \hat{g}_{h+1}}(x_h)] \right| \le &~ (H-h)(2\epsilon _V + \sqrt{4M\epsilon _V + 2\epsilon _L} + 8\tau ) + 2\epsilon _V.Next, we work with the second term in Equation (REF ):&~ \left|\mathbf {E}_{s} [\hat{g}_h(x_h)] - \mathbf {E}_{s}[g_{\h...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.00961332768201828, 0.0048791454173624516, 0.0053331078961491585, -0.03341775760054588, 0.00210005440749228, -0.000639934791252017, 0.012947473675012589, 0.02085939422249794, 0.003015609225258231, 0.014336065389215946, -0.02081361785531044, 0.009704883210361004, 0.013183992356061935, -0....
a980f6d3621afef19a908c7b0483f1026f5b39ef
subsection
73
106
Inductive case:
According to the algorithm, this only happens when the MMD test suggests that the data set \tilde{D} drawn from s looks very similar to a previous data set D \in ^{\textrm {val}}_h, which corresponds to some s^{\prime } \in ^{\textrm {val}}_h. So,&~ |\mathbf {E}_{s}[\hat{g}_h(x_h)] - \mathbf {E}_{s}[g^\star (x_h)]| \\ ...
{ "cite_spans": [] }
1803.00606
On Oracle-Efficient PAC RL with Rich Observations
[ "Christoph Dann", "Nan Jiang", "Akshay Krishnamurthy", "Alekh Agarwal", "John Langford", "Robert E. Schapire" ]
[ "cs.LG", "stat.ML" ]
2,018
en
Computer Science
[ -0.005875302944332361, 0.0023081547114998102, -0.03845652937889099, -0.018602583557367325, -0.009377594105899334, -0.003361131064593792, 0.0045094857923686504, 0.026950091123580933, 0.0017368387198075652, 0.037785064429044724, -0.027774158865213394, -0.02638545259833336, 0.007573036476969719...