text
stringlengths
16
3.88k
source
stringlengths
60
201
18.156 Lecture Notes Febrary 17, 2015 trans. Jane Wang The main goal of this lecture is to prove Korn’s inequality, which as we recall is as follows: Theorem 1 (Korn’s Inequality). If u ∈ C2 comp(Rn), and ∆u = f , then [∂i∂ju]Cα ≤ C(n, α)[∆u]Cα. First, let us recall the progress that we made last time. To start, we hav...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
. 1. f supported between x1 and x2. 1 2. f supported over x1. Used that |K(x)| (cid:46) |x|−n. 3. f supported on B3d(x1), and (cid:15) < d. Note that as opposed to can be (cid:29) dα. the previous examples, |T(cid:15)f (x1)| Used that (cid:82) Sr K(cid:15)(x) = 0 for all r, (cid:15). For this case, we will use the fol...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
dy (cid:12) (cid:12) (cid:12) (cid:12) (f (y) − A)K(cid:15)(x1 − y) dy − ( f (y) − B )K(cid:15)(x2 − y) dy N1 + (cid:90) N c 1 (f (y) − C)K(cid:15)(x1 − y) dy − (cid:90) N c 2 (f (y) − D K(cid:15)(x2 ) − y) dy (cid:12) (cid:12) (cid:12) (cid:90) N2 Let us denote the four integran may be any constants since (cid:82) Sr ...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
first two terms will behave like example 2 and the last two terms will b The third term will behave like example 3 and is the most interesting, so let us work through that bound. ve like example eha (cid:12) (cid:12) (cid:12) (cid:90) F I3 − I4 (cid:12) (cid:12) = (cid:12) ≤ ≤ (cid:90) F (cid:90) (cid:12) (cid:12) (cid:...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
15)f (x2)| + δij f (x1) | − | f (x2) . (cid:15) →0+ 1 n Eventually, (cid:15) < |x1 − x2|/10 and we can apply theorem 3 to the first term. The second term is bounded by [f ]Cα · |x1 − x2|α. To prove Korn’s inequality, we use the mollifier trick to show that we only need that u has two derivatives. Proof of Korn’s inequali...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
u(x1 − y) − ∆u(x2 − y))ϕ(cid:15)(y) dy (cid:12) (cid:12) (cid:12) ≤ [∆u]Cα|x1 − x α | 2 (cid:90) ϕ(cid:15)(y) dy. Our next goal will be to prove the Schauder Inequality. Recall that Korn’s inequality and the first homework allowed us to prove the following lemma. Lemma 6. If |aij(x) − δij| < (cid:15)(α, n) for all i, j,...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
)). Then, (cid:107)u(cid:107)C2,α(B1/2) (cid:46) max (cid:107)u(cid:107)C2,α(B(x i,r(i))) (cid:46) max u C (B(xi,r)) (cid:107) (cid:107) 2 ≤ (cid:107) (cid:107) u C2(B1). i i 5 MIT OpenCourseWare http://ocw.mit.edu 18.156 Differential Analysis II: Partial Differential Equations and Fourier Analysis Spring 2016 For inf...
https://ocw.mit.edu/courses/18-156-differential-analysis-ii-partial-differential-equations-and-fourier-analysis-spring-2016/2a167573b1412e48588f173774f6e158_MIT18_156S16_lec5.pdf
6.034 Artificial Intelligence. Copyright © 2004 by Massachusetts Institute of Technology. 6.034 Notes: Section 7.1 Slide 7.1.1 We have been using this simulated bankruptcy data set to illustrate the different learning algorithms that operate on continuous data. Recall that R is supposed to be the ratio of earnings ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
as simple as this class is, in general, there will be many possible linear separators to choose from. Also, note that, once again, this decision boundary disagrees with that drawn by the previous algorithms. So, there will be some data sets where a linear separator is ideally suited to the data. For example, it tur...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
1 values, the components of an n-dimensional coefficient vector w and a scalar value b. These n+1 values are what will be learned from the data. The x will be some point in the feature space. We will be using dot product notation for compactness and to highlight the geometric interpretation of this equation (more o...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
linear separators in n dimensions. In two dimensions, such a linear separator is refered to as a "line". In three dimensions, it is called a "plane". These are familiar words. What do we call it in higher dimensions? The usual terminology is hyperplane. I know that sounds like some type of fast aircraft, but that's ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
x are perpendicular), the cosine is equal to 0 and the distance is precisely b as we expect. Slide 7.1.14 This distance measure from the hyperplane is signed. It is zero for points on the hyperplane, it is positive for points in the side of the space towards which the normal vector points, and negative for points ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
. Then we repeat the inner loop until all the points are correctly classified using the current weight vector. The inner loop is to consider each point. If the point's margin is positive then it is correctly classified and we do nothing. Otherwise, if it is negative or zero, we have a mistake and we want to change t...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
. Here it took 49 iterations through the data (the outer loop) for the algorithm to stop. The hypothesis at the end of each loop is shown here. Recall that the first element of the weight vector is actually the offset. So, the normal vector to the separating hyperplane is [0.94 0.4] and the offset is -2.2 (recall th...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
of this strategy to multiple input variables is based on the generalization of the notion of slope, which is the gradient of the function. The gradient is the vector of first (partial) derivatives of the function with respect to each of the input variables. The gradient vector points in the direction of steepest inc...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
discard the points after using them, counting on more arriving later. Another way of thinking about the relationship of these algorithms is that the on-line version is using a (randomized) approximation to the gradient at each point. It is randomized in the sense that rather than taking a step based on the true grad...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
to 0s, we can write the final weight vector in terms of these counts and the input data (as well as the rate constant). Slide 7.2.13 Since the rate constant does not change the separator we can simply assume that it is 1 and ignore it. Now, we can substitute this form of the weights in the classifier and we get the...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
takes place by adjusting the weights in the network so that the desired output is produced whenever a sample in the input data set is presented. Slide 7.3.2 We start by looking at a simpler kind of "neural-like" unit called a perceptron. This is where the perceptron algorithm that we saw earlier came from. Percept...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
seen in our treatment of SVMs how the "kernel trick" can be used to generalize a perceptron-like classifier to produce arbitrary boundaries, basically by mapping into a high- dimensional space of non-linear mappings of the input features. Slide 7.3.6 We will now explore a different approach (although later we will a...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
feature values (1,1)) is in the half space that the normal points into. This is the only point with a positive distance and thus a one output from the perceptron unit. The other points have negative distance and produce a zero output. This is shown in the shaded column in the table. Slide 7.3.9 Looking at the secon...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
it doesn't matter how far a point is from the decision boundary, you will still get a 0 or a 1. We need a smooth output (as a function of changes in the network weights) if we're to do gradient descent. Slide 7.3.13 Eventually people realized that if one "softened" the thresholds, one could get information as to w...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
see what that means. The output of a multi-layer net of sigmoid units is a function of two vectors, the inputs (x) and the weights (w). An example of what that function looks like for a simple net is shown along the bottom, where s() is whatever output function we are using, for example, the logistic function we sa...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
single training point. Thus, we will be neglecting the sum over the training points in the real gradient. As we saw in the last slide, we will need the gradient of the unit's output with respect to the weights, that is, the vector of changes in the output due to a change in each of the weights. The output (y) of a ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
in these gradients, not surprisingly. Here we see that this derivative has a very simple form when expressed in terms of the output of the sigmoid. Then, it is just the output times 1 minus the output. We will use this fact liberally later. Slide 7.3.22 Now, what happens if the input to our unit is not a direct inp...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
at the output (y-y^i) times the change in the output which is produced by the change in the weight (dy/dw). Slide 7.3.27 Let's pick weight w13, that weights the output of unit 1 (y1) coming into the output unit (unit 3). What is the change in the output y3 as a result of a small change in w13? Intuitively, we shoul...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
strategy for computing the error gradient. Slide 7.3.29 The cases we have seen so far are not completely general in that there has been only one path through the network for the change in a weight to affect the output. It is easy to see that in more general networks there will be multiple such paths, such as shown ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
and moving backward through the network we can compute all the deltas for every unit in the network in one pass (once we've computed all the y's and z's during a forward pass). It is this property that has led to the name of this algorithm, namely backpropagation. It is important to remember that this is still the ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
randomness into the gradient descent. Formally, gradient descent on an error function defined as the sum of the errors over all the input instances should be the sum of the gradients over all the instances. However, backprop is typically implemented as shown here, making the weight change based on each feature vecto...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
derive the gradient of the network and a backprop-like computation can be used to do that. 6.034 Artificial Intelligence. Copyright © 2004 by Massachusetts Institute of Technology. 6.034 Notes: Section 7.4 Slide 7.4.1 Now that we have looked at the basic mathematical techniques for minimizing the training error o...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
training data. Instead, we can use the performance on the validation set as a way of deciding when to stop; we want to stop when we get best performance on the validation set. This is likely to lead to better generalization. We will look at this in more detail momentarily. This type of "early termination" keeps the...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
(drawn from Gaussian distributions with different centers). An additional 25 instances each (drawn from the same distributions) have been reserved as a test set. As you can see, the point distributions overlap and therefore the net cannot fully separate the data. The red region represents the area where the net's ou...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
that are just as unlikely to generalize to new data. For K-nearest-neighbors, on this type of data one would want to use a value of K greater than 1. For decision trees one would want to prune back the tree somewhat. These decisions could be based on the performance on a held out validation set. Similarly, for neur...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
other hand, the need to avoid overstepping the minimum and possibly getting into oscillations because of a too-large learning rate. One approach to balancing these is to effectively adjust the learning rate based on history. One of the original approaches for this is to use a momentum term in backprop. Here is the ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
is best to keep the range of the inputs in that range as well. Simple normalization (subtract the mean and divide by the standard deviation) will almost do that and is adequate for most purposes. Slide 7.4.17 Another issue has to do with the representation of discrete data (also known as "categorical" data). You c...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
or stops changing significantly. These outcomes generally happen long before we run the risk of the weights becoming infinite. 6.034 Artificial Intelligence. Copyright © 2004 by Massachusetts Institute of Technology. Slide 7.4.20 Neural nets can also do regression, that is, produce an output which is a real numbe...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
in the road got you off center? It is important that ALVINN be able to recover and steer the vehicle back to the center. The researchers considered having the vehicle drive in a wobbly path during training, but that posed the danger of having the system learn to drive that way. They came up with a clever solution. ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a180cde9d8748aa97cbf51950ecb096_ch7_mach3.pdf
6.034 Artificial Intelligence. Copyright © 2004 by Massachusetts Institute of Technology. 6.034 Notes: Section 3.4 Slide 3.4.1 In this section, we will look at some of the basic approaches for building programs that play two- person games such as tic-tac-toe, checkers and chess. Much of the work in this area has be...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
complex game. 6.034 Artificial Intelligence. Copyright © 2004 by Massachusetts Institute of Technology. Slide 3.4.4 Here's a little piece of the game tree for Tic-Tac-Toe, starting from an empty board. Note that even for this trivial game, the search tree is quite big. Slide 3.4.5 A crucial component of any game...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
The player who is building the tree is trying to maximize their score. However, we assume that the opponent (who values board positions using the same static evaluation function) is trying to minimize the score (or think of this is as maximizing the negative of the score). So, each layer of the tree can be classifie...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
world champion with a ranking of about 2900. At some level, this is a depressing picture, since it seems to suggest that brute-force search is all that matters. Slide 3.4.10 And Deep Blue is brute indeed... It had 256 specialized chess processors coupled into a 32 node supercomputer. It examined around 30 billion ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
Here's some pseudo-code that captures this idea. We start out with the range of possible scores (as defined by alpha and beta) going from minus infinity to plus infinity. Alpha represents the lower bound and beta the upper bound. We call Max-Value with the current board state. If we are at a leaf, we return the stat...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
3.4.20 The calling Max-Value now sets alpha to this value, since it is bigger than minus infinity. Note that the range of [alpha beta] says that the score will be greater or equal to 2 (and less than infinity). Slide 3.4.21 Max-Value now calls Min-Value with the updated range of [alpha beta]. Slide 3.4.22 Min-Val...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
as deep! We already saw the enormous impact of deeper search on performance. So, this one simple algorithm can almost double the search depth. Now, this analysis is optimistic, since if we could order moves perfectly, we would not need alpha- beta. But, in practice, performance is close to the optimistic limit. Sli...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
(killer moves) and try them first when considering subsequent moves at that level. Imagine a position with white to move. After white's first move we go into the next recursion of Alpha-Beta and find a move K for black which causes a beta cutoff for black. The reasoning is then that move K is a good move for black, ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
alpha and beta. Instead of starting with alpha and beta at minus and plus infinity, we can start them in a small window around the values found in the previous search. This will help us cutoff more irrelevant moves early. In fact, it is often useful to start with the tightest possible window, something like [alpha, ...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
seems not to be susceptible to search- based methods that work well in chess. Go players seem to rely more on a complex understanding of spatial patterns, which might argue for a method that is based more strongly on a good evaluation function than on brute-force search. Slide 3.4.33 There are a few observations ab...
https://ocw.mit.edu/courses/6-034-artificial-intelligence-spring-2005/2a2f347357dc1e683ece44aec2f81e5e_ch3_csp_games2.pdf
Chapter 2 Abelian Gauge Symmetry As I have already mentioned, local gauge symmetry is the major new principle, beyond the generic implementation of special relativity and quantum mechanics in quantum field theory, which arises in formulating the standard model. It is, however, a rather abstract concept, and one whose...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
Gauge symmetry is a family of functional transformations among the potentials that leaves the field strength unchanged. The charge and current distributions, which provide the source terms in the Maxwell equations, are under gauge transformations. In the general, classical continuum form the Maxwell equations the ph...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
here on call potentials – we add to the action a term corresponding to the Lagrangian Sint. = q Aµdxµ − � The momentum is now, with L Lint. = dxj dt . qA0 + qAj − Lf ree + Lint., ≡ ∂L ∂ ˙xj = m pj = vj √1 v2 − + qAj dpj dt = d dt ( mvj √1 �v2 − ) + q ∂Aj ∂xk dxj dt + q ∂Aj ∂t . and Si...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
0 ∂xj − ∂Aj ∂t ≡ − Bj �jlm ∂Al ∂xm ≡ 5 (2.11) (2.12) Identifying � E and � B as electric and magnetic field strengths, we thereby arrive at the Lorentz force law for a particle of mass m, charge q. Note that with this identification two of the Maxwell equations, viz. ∂Bj ∂xj ∂Bj ∂t = 0 = 0 �jlm ∂El ...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
motion unchanged, it must leave � B unchanged, as of course one can verify directly. E and � � Clearly, the requirement that the world-line of a charged particle should have no ends is closely related to the conservation of charge. More generally, the necessary and sufficient condition for an action Sint. = − � ...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
+ qA0. (2.21) � − The appearance of square root here leads to difficulties in quantization. In order to implement the commutation relations, or (more heuristically) wave-particle duality, one would like to make the substitution � p odinger wave equation Hψ = i ∂ψ . But that substitution renders the Hamiltonian, wi...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
.23) i ∂ψ ∂t i∂� ( − q � A)2 absorbing a factor e− imt into ψ.) 7 For the gauge transformation A� = Aµ + ∂µχ on the potentials to leave the µ Schr¨odinger equation invariant, it must be accomplished by the transformation ψ� = e− iqχψ (2.24) on the charged fields. Note that the value of charge q appears expl...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
2 ψp1 2 · · · p1q1 + p2q2 + = 0 · · · (2.25) (2.26) for the term destroys p1 particles charge q1, p2 particles charge q2, and so forth (or creates an equivalent number of anti-particles). Putting it differently, the necessary and sufficient condition for charge conservation is invariance under the symmetry ψ� =...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
modify the derivatives. (2.28) A suitable modification is suggested by our earlier discussion of the point particle Hamiltonian. We define the covariant derivative operate Dµ to act on a field ψn of charge qn according to Then if Aµ transforms as in Equation 2.15, we have Dµψn = (∂µ + iqnAµ)ψn. (2.29) 2.1. GAUGE...
https://ocw.mit.edu/courses/8-325-relativistic-quantum-field-theory-iii-spring-2003/2a43685db9124bc04925067ae33c6f23_chap2.pdf
MIT 2.852 Manufacturing Systems Analysis Lectures 18–19 Loops Stanley B. Gershwin Spring, 2007 Copyright c�2007 Stanley B. Gershwin. Problem Statement B1 M2 B2 M3 B3 M1 M4 B6 M6 B5 M5 B4 • Finite buffers (0 � ni(t) � Ni). • Single closed loop – fixed population ( • Focus is on the Buzacott model (deter...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
d o r p 0.89 0.885 0.88 0.875 0.87 0.865 0.86 0 N2=10 N2=15 N2=20 N2=30 N2=40 10 20 30 40 50 60 population Copyright �2007 Stanley B. Gershwin. c 5 Expected population method • Treat the loop as a line in which the first machine and the last are the same. • In the resulting decomposition, on...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
uses E(i) = E(i + 1) for i = 1, ..., k. The first k − 1 are E(1) = E(2), E(2) = E(3), ..., E(k − 1) = E(k). But this implies E(k) = E(1), which is the same as the kth equation. Copyright �2007 Stanley B. Gershwin. c 7 Expected population method Therefore, we need one more equation. We can use n¯i = N � i ...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
machines. In a line, every downstream machine could block a given machine, and every upstream machine could starve it. In a loop, blocking and starvation are more complicated. Copyright c �2007 Stanley B. Gershwin. 11 Loop Behavior Ranges B1 M2 B2 M3 B3 M1 M4 B6 M6 B5 M5 B4 • The range of blocking of...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
part of the line. • The range of starvation of a machine in a line is the entire upstream part of the line. Copyright c �2007 Stanley B. Gershwin. 15 Loop Behavior Ranges Line In an acyclic network, if Mj is downstream of Mi, then the range of blocking of Mj is a subset of the range of blocking of Mi. M i M...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
c 18 Loop Behavior Range of blocking of M 1 Range of blocking of M 2 10 B1 7 B6 M1 10 10 B 2 0 B5 M 2 M6 Ranges Difficulty for decomposition 10 B3 0 B4 M4 M3 M5 M5 can block M2. Therefore the parameters of M5 should directly affect the parameters of Md(1) in a decomposition. However, M5 cannot ...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
machine of its two-machine line. • Similarly for upstream modes. Copyright �2007 Stanley B. Gershwin. c 22 Multiple Failure Mode Line Decomposition 1,2 3 4 5,6,7 8 9,10 1,2, 3,4 5,6,7, 8,9,10 • The downstream failure modes appear to the observer after propagation through blockage . • The upstream fail...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
line. • Local modes: the probability of failure into a local mode is the same as the probability of failure in that mode of the real machine. Copyright �2007 Stanley B. Gershwin. c 26 Multiple Failure Mode Line Decomposition Line Decomposition 1,2 3 4 5,6,7 8 9,10 • Remote modes: i is the building block num...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
Copyright �2007 Stanley B. Gershwin. c 28 Multiple Failure Mode Line Decomposition Extension to Loops B1 M 2 B 2 M3 B3 M1 M4 B6 M6 B5 M5 B4 • Use the multiple-mode decomposition, but adjust the ranges of blocking and starvation accordingly. • However, this does not take into account the local inform...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
. • Consequently, this makes the two-machine line very complicated. Copyright c �2007 Stanley B. Gershwin. 32 Transformation • Purpose: to avoid the complexities caused by thresholds. • Idea: Wherever there is a threshold in a buffer, break up the buffer into smaller buffers separated by perfectly reliable mac...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
has none. • Note: The number of thresh­ olds equals the number of ma­ chines. 5 35 Transformation B 1 20 M2 B 2 3 * M1 M1 M3 15 5 B 4 M4 B 3 1M * M4 * M3 * M2 M2 M3 M4 • Break up each buffer into a sequence of buffers of size 1 and reliable machines. • Count backwards from each real mac...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
were as high as 10%. � Six-machine cases: mean throughput error 1.1% with a maximum of 2.7%; average buffer level error 5% with a maximum of 21%. � Ten-machine cases: mean throughput error 1.4% with a maximum of 4%; average buffer level error 6% with a maximum of 44%. Copyright �2007 Stanley B. Gershwin. c 40 Numeri...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 • Production rate vs. r1. • Usual saturating graph. Copyright �2007 Stanley B. Gershwin. c 44 Numerical Results B1 M2 B2 * M1 B4 M4 B3 10 9 8 7 6 5 4 3 2 1 0 0 M3 Behavior b1 average b2 average b3 average b4 average 0.1 0.2 0.3 0.4 0.5 ...
https://ocw.mit.edu/courses/2-852-manufacturing-systems-analysis-spring-2010/2a492e1e011aa53ffe87fefef19d6461_MIT2_852S10_loops.pdf
Practical multitone architectures Lecture 4 Vladimir Stojanović 6.973 Communication System Design – Spring 2006 Massachusetts Institute of Technology Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006. MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Te...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
Efficientizing example ‰ 1+0.9D-1 channel (Pe=10-6, gap=8.8dB, PAM/QAM) „ PAM and single-sideband „ QAM bn b +1n Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006. MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Month...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
Bit is moved from channel n to m ‰ Tightly coupled with channel and noise estimation „ Will cover in later lectures Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006. MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Mon...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY]. 6.973 Communication System Design 8 Modal modulation ‰ Transmission with eigen-functions r(t) = h(t) * h*(-t) (-TH / 2, TH / 2) . rn jn (t) = T / 2 -T / 2 (cid:242) r(t - t)jn(t)dt n = 1,..., , t [-(T - TH) ...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
of any autocorrelation function is unique „ This set determines the performance of MM through SNR „ Eigen-functions are not unique „ is also a valid eigen-function for inf symbol period „ Corresponding eigenvalues are R(2πn/T) „ No ISI on any tone since symbol period is infinite „ Each tone is AWGN channel „ SB...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
p1 p0 p0 0 0 0 py py-1 0 p0 0 py p1 0 0 0 py xN-1 x0 x-1 x-y + nN-1 nN-1 n0 Figure by MIT OpenCourseWare. Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006. MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY]. 6....
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
-varying channels „ Optimize bn and En per sub-channel ‰ OFDM uses same channel partitioning as DMT „ But uses same bn and En on all channels „ Used on one-way broadcast channels ‰ Forms of vector coding with added restrictions „ In vector coding, M,F channel dependent „ Make the channel circular and make M,F ch...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
DD Month YYYY]. 6.973 Communication System Design 17 DMT/OFDM implementation X0 X1 . . . XN-2 XN-1 1 N+v G = T 1 N+v G = T IDFT parallel to serial & cyclic prefix insert x(t) n(t) D A C j(t) (LPF) h(t) + j (-t) (LPF) y(t) A D C Forces P to be cyclic matrix serial to parallel & cyclic prefix remover DFT Y0 Y1 . . . Y...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
CP lower than VC (8.1dB) „ For N=16 quickly reaches max of 8.8dB Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006. MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY]. 6.973 Communication System Design 21 C...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
1.5 Mbps Split ADSL Down Down frequency z H M 1 1 . Figure by MIT OpenCourseWare. „ Hermitian symmetry creates real signal transmitted from 0-1.1MHz „ First 2-3 tones near DC not used – avoid interference with voice „ Tone 256 also not used, 64 reserved for pilot ‰ Nup=32, CP=5, each symbol 2*32+5=69 samples „ Exac...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
3/2 3 9/2 bn 1/4 3/8 1/2 3/4 1 3/4 3/2 9/4 b 24 36 48 72 96 144 192 216 ‰ Broadcast channel – can’t optimize bit allocation Figure by MIT OpenCourseWare. „ FCC demands flat spectrum so no energy-allocation „ The only knob is data rate selection Cite as: Vladimir Stojanovic, course materials for 6.973 Communication S...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
r l e a v e r M a p p e r D e m a p p e r P i l o t i n s e r t i o n C h a n n e l e s t i m a t o r F F T / I F F T C y c l i c p r e f i x S y n c h r o n z e r i i W n d o w n g i R e m o v e p r e f i x D A C A G C & A D C U p c o n v e r t L N A & D o w n c o n v e r t Cite as: Vladimir Stojanovic, course materia...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
Scrambling ‰ Need to randomize incoming data ‰ Enables a number of tracking algorithms in the receiver ‰ Provides flat spectrum in the given band pseudo-random bit sequence (prbs) generator What is the period of this pseudo-random sequence? Cite as: Vladimir Stojanovic, course materials for 6.973 Communication Sys...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
B4 3 A5 A A 7 6 B7 B B 6 5 A8 B 8 Stolen Bit Bit Inserted Data Output Data B A2 A A 0 1 A A 3 4 B B B B B B B B B 4 8 AA 6 5 A A 7 0 1 3 5 2 6 7 8 Inserted Dummy Bit g1=1718 Decoded Data y 0 y 1 y 2 y 3 y 4 y 5 y 6 y 7 y 8 ‰ 64-state (constraint length K=7) code ‰ Viterbi algorithm applied in the decoder Figure by...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
CourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY]. 6.973 Communication System Design 32 Spectral mask ‰ Cannot use last 5 tones on each side ‰ Does not use extra windowing Power Spectral Density (dB) Transmit Spectrum Mask (not to scale) Typical Signal Spectru...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
t8 t9t10 GI2 T1 T2 GI SIGNAL GI Data 1 GI Data 2 Signal Detect, AGC, Diversity Selection Coarse Freq. Offset Estimation Timing Synchronize Channel and Fine Frequency Offset Estimation RATE LENGTH SERVICE + DATA DATA Figure by MIT OpenCourseWare. to timing control symbol timing adjust training symbol pilots angle adjust...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
H(i-D,k) = Y(i-D,k) Mapper Interleaver Conv. Encoder X(i-D,k) X(i-D,k) Figure by MIT OpenCourseWare. Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006. MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY]. 6.973 C...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
Y.-H. Wang and R. Yu "An integrated 802.11a baseband and MAC processor," Solid-State Circuits Conference, 2002. Digest of Technical Papers. ISSCC. 2002 IEEE International vol. 1, no. SN -, pp. 126- 451 vol.1, 2002. [4] E. Grass, K. Tittelbach-Helmrich, U. Jagdhold, A. Troya, G. Lippert, O. Kruger, J. Lehmann, K. Mah...
https://ocw.mit.edu/courses/6-973-communication-system-design-spring-2006/2a8950c0b8a77687bae83a70ff459b08_lecture_4.pdf
Massachusetts Institute of Technology Department of Electrical Engineering and Computer Science 6.245: MULTIVARIABLE CONTROL SYSTEMS by A. Megretski Fundamentals of Model Order Reduction1 This lecture introduces basic principles of model order reduction for LTI systems, which is about finding good low order approx...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
ˆ = Sˆk of complexity not larger than a given threshold k, such that the distance between Sˆ and a given ”complex” system S is as small as possible. Alterna­ tively, a maximal admissible distance between S and Sˆ can be given, in which case the complexity k of Sˆ is to be minimized. As is suggested by the experience...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
, W are given stable transfer matrices (W −1 is also assumed to be stable), and ≡�≡� denotes H-Infinity norm (L2 gain) of a stable system �. As a result of model order reduction, G can be represented as a series connection of a lower order “nominal plant” ˆG and a bounded uncertainty (see Figure 8.2). In most cases, ...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
problem reducible to solving a system of linear equations. If ≡ · ≡ = ≡ · ≡� is the H-infinity norm, the optimization is reduced to solving a system of Linear Matrix Inequalities (LMI), a special class of convex optimization problems solvable by an efficient algorithm, to be discussed later in the course. While the te...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
norm optimal model reduction problem yields a lower bound for the minimum in the H- Infinity norm optimal model reduction. Moreover, H-Infinity norm of model reduction error associated with a Hankel norm optimal reduced model is typically close to this lower bound. Thus, Hankel norm optimal reduced model can serve well...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
q are polynomials of order m. ˆ One popular approach is moments matching. In the simplest form of moments match- ing, an m-th order approximation G(s) = p(s)/q(s) (where p, q are polynomials of order m) of a SISO transfer function G(s) is defined by matching the first 2m + 1 coefficients of a Taylor expansion ˆ G(s) = ...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
· · · + q1z + q0, with qm ∞= 0, such that qmyk+m + qm−1yk+m−1 + · · · + q1yk+1 + q0yk = 0 for all k > 0. The idea is to define the denominator of the reduced model G(s) in terms of a polynomial q = q(z) which minimizes the sum of squares of ˆ ek = qmyk+m + qm−1yk+m−1 + · · · + q1yk+1 + q0yk subject to a normalizat...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
The “convolution operator” f ∈� y = g � f associated with a LTI system with impulse response g = g(t) has infinite rank whenever g is not identically equal to zero. However, with every LTI system of finite order, it is possible to associate a meaningful and repre­ sentative linear transformation of finite rank. This tra...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
the complex conjugates of the corresponding entries of G(jρ)) and continuity (i.e. G(jρ) converges to G(jρ0) as ρ � ρ0 for all ρ0 ≤ R and for ρ0 = ∗). Note that a rational function G = G(s) with real coefficients satisfies this condition if and only if it is proper and has no poles on the imaginary axis. Let Lk 2 (−∗...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
gain readily defined: the rank of HG is the maximal number of linearly independent 2 (−∗, 0) ∈� Lm 7 outputs (could be plus infinity), the L2 gain is the minimal upper bound for the L2 norm of HGf subject to the L2 norm of f being fixed at one. Remember that the order of a rational transfer matrix G is defined as t...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
the L2 norm of g0 is not larger than ≡G≡�≡f ≡2. On the other hand, ≡g≡2 ∪ ≡g≡2. To show that the rank of HG equals the order of the stable part of G, note first that the unstable part of G does not have any effect of HG at all. Then, for a stable G, the map of f into the inverse Fourier transform g0 of ˜g0 = Gf˜ is a c...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
efficient solution. 8 Let m M = � ur λr vr ⎦ r=1 be a singular value decomposition of M , which means that the families {ur }m and {vr }m are orthonormal, and {λr }m is a monotonically decreasing sequence of positive r=1 numbers, i.e. r=1 r=1 � uiur = vivr = � 1, 0, i = r, i ∞= r, ⎤ λ1 → λ2 → · · · → λ...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
λr = |M vr | when transformed with M . Vector ur is defined by M vr = λr ur . Another useful interpretation of singular vectors vr , ur and singular numbers λr is by the eigenvalue decompositions (M �M )vr = λ2 vr , r (M M �)sur = λ2 ur . r An approximation M = Mk of rank less than k which minimizes λmax(M − ˆ ˆ ˆ...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf
numbers λr = λr (M ) will be used: if the first k right singular vectors v1, . . . , vk of M can be defined, but the supremum � = sup{|M v| : |v| = 1, v orthogonal to v1, . . . , vk } is not achievable, then λr (M ) = � for all r > k. 8.2.4 SVD of a Hankel operator Any causal stable LTI system with a state-space mod...
https://ocw.mit.edu/courses/6-245-multivariable-control-systems-spring-2004/2abdecbaa80cd2d06ffa931ca7ca34c3_lec8_6245_2004.pdf