dm2 data for 3sv4z and 5v6
Browse files- 3sv4z/3sv4z_data_note.txt +11 -0
- 3sv4z/ippo_3sv4z.zip +3 -0
- 3sv4z/ippo_3sv4z_tb_logs.zip +3 -0
- 3sv4z/qmix_3sv4z.zip +3 -0
- 5v6/5v6_data_note.txt +13 -0
- 5v6/ippo_5v6_tb_logs.zip +3 -0
- 5v6/qmix_5v6.zip +3 -0
3sv4z/3sv4z_data_note.txt
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Data used for 3sv4z experiments:
|
| 2 |
+
|
| 3 |
+
QMIX demonstrations: use path
|
| 4 |
+
"qmix_3sv4z/qmix_sc2_saved-data=3003-eval-only-eps=0.05__seed=111111_07-31-14-56-08/"
|
| 5 |
+
this data was sampled from the expert QMIX policy with random action epsilon=0.05. we found that having
|
| 6 |
+
some stochasticity in the expert data was necessary for IPPO to learn well.
|
| 7 |
+
the path "qmix_3sv4z/qmix_sc2_saved-data=3003-eval-only__seed=111111_07-27-17-04-09/_eval" contains data
|
| 8 |
+
that was sampled with epsilon=0.0
|
| 9 |
+
|
| 10 |
+
IPPO demonstrations: use path "ippo_3sv4z/ippo_sc2_baseline_seed=119527_01-23-20-41-57/". as indicated in the paper,
|
| 11 |
+
we used timesteps corresponding to 5M and 7M steps.
|
3sv4z/ippo_3sv4z.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e92e17cc11af7187cc254741259e9b61427736e5d521545e32e968f46f273daf
|
| 3 |
+
size 1313716793
|
3sv4z/ippo_3sv4z_tb_logs.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3f6c8cfda65914913c9d35be4c66d788cc8a10b8ce3342c6601d48cbe93653af
|
| 3 |
+
size 5810870
|
3sv4z/qmix_3sv4z.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:75c4b153d196354ca580c2ecdcf0c1626865654a3c4bdc945b82efaf2c217044
|
| 3 |
+
size 242299399
|
5v6/5v6_data_note.txt
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
QMIX demos:
|
| 2 |
+
use path qmix_5v6/qmix_sc2_saved-data=5005-eval-only-eps=0.025__seed=111111_07-31-14-41-26/_eval
|
| 3 |
+
we use data sampled with epsilon=0.025
|
| 4 |
+
|
| 5 |
+
iPPO demos:
|
| 6 |
+
use path ippo_5v6/agents_batches/ippo_sc2_saved-batchsize=5005_seed=112358_03-06-21-12-41/
|
| 7 |
+
as indicated in the paper, we use timesteps 5M and 7M.
|
| 8 |
+
|
| 9 |
+
the other seeds (with the same name pattern) are for the ablation experiment where we use a different seed of demonstration data
|
| 10 |
+
for each agent in the team.
|
| 11 |
+
|
| 12 |
+
the path ippo_5v6/agents_batches/ippo_sc2_saved-batchsize=5000-ts=<TS>-concurrent-demos-diff-policies-DATAONLY_seed=92814_01-17-22-42-49/_eval/
|
| 13 |
+
is for the ablation where we use demo data created by running policies trained w/different seeds in the same environment.
|
5v6/ippo_5v6_tb_logs.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9224207aefda223e707d971f6c1d34311b3b91fe36a6a6c964892f6e78ad6b32
|
| 3 |
+
size 19776169
|
5v6/qmix_5v6.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e129163806e2c1aabe6eb03226068a9344dd517c2beefc6725e97acb5857aaa8
|
| 3 |
+
size 467556181
|