Code for "Joint Policy Search for Collaborative Multi-agent Imperfect Information Games". Arxiv link. The paper is published in NeurIPS 2020.
The project aims to find better equilibrium in multi-agent collaborative games with imperfect information by improving the policies of multiple agents simultaneously. This helps escape local equilibrium where unlaterial improvement of one player's policy is not helpful. To achieve that, we developed a novel value decomposition technique that decomposes the expected value changes into information sets where the policy differs, and search over candidate information sets via depth-first search. Each update can be proven not to degrade the performance in the tabular cases.
We open source the code to reproduce our results in simple games (see Def. 1-3 in the paper). The dataset and pre-trained Bridge model will be released later.
@inproceedings{tian2020jps,
title={Joint Policy Search for Multi-agent Collaboration with Imperfect Information},
author={Yuandong Tian and Qucheng Gong and Tina Jiang},
booktitle={NeurIPS},
year={2020}
}
Thanks Xiaomeng Yang (@xiaomengy) for substantial code refactor after the paper acceptance.
Compiled with Linux and GCC 12.1 Cuda 10.1 PyTorch 1.7.0
- Note that 1.7.1+ do not work due to some conflict with pybind11. The current submodule
third_party/pybind11is set to be at commita1b71df. - You may try using later version of pybind11 and newer version of PyTorch (not tested yet).
| Model | Description |
|---|---|
| agent-baseline-15-189.pth | Baseline model1 |
| agent-baseline-0704-26-199.pth | Baseline model2 |
| agent-2day-5-250.pth | JPS models (0.38 IMPs/b against WBridge5) |
| agent-2day-8-169.pth | JPS models (0.44 IMPs/b against WBridge5) |
| agent-1610-0.63.pth | JPS models (0.63 IMPs/b against WBridge5) |
conda create --name bridge --file requirement.txt
conda activate bridge
git submodule update --init --recursive
cd third_party/pybind11
git checkout a1b71df
cd ../../
cmake ..
make
cp ./*.so ./python
cp ./rela/*.so ./python
cd python
Please download the dataset here. The dataset contains:
- The training set (
dda.db) contains 2.5M situations, - The evaluation set (
test.db) contains 50K situations. - The VS WBridge5 set (
vs_wb5.db) contains 1000 situations.
Please untar it to ./bridge_data/ folder.
To check the record, you can install sqlite3 and run the following SQLs:
- Run
select count(*) from records;to check the number of records - Run
select * from records limit 1;to check the first entry.
Here is one example:
$ sqlite3
SQLite version 3.31.1 2020-01-27 19:55:54
Enter ".help" for usage hints.
Connected to a transient in-memory database.
Use ".open FILENAME" to reopen on a persistent database.
sqlite> .open dda.db
sqlite> select * from records limit 1;
0|{"pbn": "[Deal \"N:KT9743.AQT43.J.7 J85.9.Q6.KQJ9532 Q2.KJ765.T98.T64 A6.82.AK75432.A8\"]", "ddt": [0, 12, 0, 12, 0, 12, 0, 12, 10, 3, 10, 3, 9, 4, 9, 4, 0, 8, 0, 8]}
Note that each entry has a pre-computed double dummy table. DDS table is in the format of C (NESW), D (NESW), H (NESW), S (NESW), NT (NESW). For example, the 1D ddt table above means that
| C | D | H | S | NT | |
|---|---|---|---|---|---|
| N | 0 | 0 | 10 | 9 | 0 |
| E | 12 | 12 | 3 | 4 | 8 |
| S | 0 | 0 | 10 | 9 | 0 |
| W | 12 | 12 | 3 | 4 | 8 |
To run the evaluation, use the following command to make pre-trained models play with each other.
python main2.py num_thread=200 game=bridge env_actor_gen.params.gen_type=basic seed=1 \
trainer=selfplay actor_gen.params.batchsize=1024 epoch_len=10 eval_only=true \
method=a2c agent.params.load_model=`pwd`/../bridge_models/agent-1610-0.63.pth \
baseline=a2c baseline.agent.params.load_model=`pwd`/../bridge_models/agent-baseline-15-189.pth \
game.params.feature_version=old/single \
game.params.train_dataset=`pwd`/../bridge_data/dda.db \
game.params.test_dataset=`pwd`/../bridge_data/test.db
Here is the example output:
[2023-11-13 15:45:14,526][main2.py][INFO] - [-1] Time spent = 23.77 s
-1:eval_score_p0 [50000]: avg: 0.0318 (± 0.0012), min: -0.8750[25576], max: 0.9167[14711]
-1:eval_score_p1 [50000]: avg: -0.0318 (± 0.0012), min: -0.9167[14711], max: 0.8750[25576]
-1:eval_score_p2 [50000]: avg: 0.0318 (± 0.0012), min: -0.8750[25576], max: 0.9167[14711]
-1:eval_score_p3 [50000]: avg: -0.0318 (± 0.0012), min: -0.9167[14711], max: 0.8750[25576]
Here 0.0318 means that the JPS model agent-1610-0.63.pth is 0.0318 * 24 = 0.7632 IMPs/b better than the baseline agent-baseline-15-189.pth, since the reward score is normalzed to [-1,1] from [-24,24] in IMPs/b scale. Note that the bridge game has 4 players and the evaluation dumps the scores for each player over 50000 games of the evaluation set (test.db)
For the three JPS models, we use the old feature while for baselines we use the single feature (which is a newer type of feature). Use game.params.feature_version=old/single to specify the features used for each model.
Another example is to compete with baseline16:
python main2.py num_thread=200 game=bridge env_actor_gen.params.gen_type=basic seed=1 \
trainer=selfplay actor_gen.params.batchsize=1024 epoch_len=10 eval_only=true \
method=a2c agent.params.load_model=`pwd`/../bridge_models/agent-1610-0.63.pth \
baseline=baseline16 \
game.params.feature_version=old \
game.params.train_dataset=`pwd`/../bridge_data/dda.db \
game.params.test_dataset=`pwd`/../bridge_data/test.db
Here is the example output:
[2023-11-13 16:17:39,622][main2.py][INFO] - [-1] Time spent = 15.74 s
-1:eval_score_p0 [50000]: avg: 0.1276 (± 0.0011), min: -0.7083[39419], max: 0.9167[13675]
-1:eval_score_p1 [50000]: avg: -0.1276 (± 0.0011), min: -0.9167[13675], max: 0.7083[39419]
-1:eval_score_p2 [50000]: avg: 0.1276 (± 0.0011), min: -0.7083[39419], max: 0.9167[13675]
-1:eval_score_p3 [50000]: avg: -0.1276 (± 0.0011), min: -0.9167[13675], max: 0.7083[39419]
To train the model, run the following:
python main2.py num_thread=100 num_game_per_thread=50 game=bridge env_actor_gen.params.gen_type=basic seed=1 \
trainer=selfplay actor_gen.params.batchsize=512 method=a2c \
game.params.train_dataset=`pwd`/../bridge_data/dda.db \
game.params.test_dataset=`pwd`/../bridge_data/test.db \
agent.params.explore_ratio=0.000625
If not specified, the default baseline is baseline16. Here is some evaluation output after 3 epochs (note that performance may vary depending on your random seed and other factors, which is what happens in RL).
Finished creating 100 eval environments
[2023-11-14 10:50:41,405][main2.py][INFO] - [3] Time spent = 671.07 s
3:R [10000]: avg: -0.0003 (± 0.0001), min: -0.0187[9474], max: 0.0170[6922]
3:V [10000]: avg: -0.0003 (± 0.0000), min: -0.0071[5170], max: 0.0063[9612]
3:adv_err [10000]: avg: 0.0113 (± 0.0000), min: 0.0068[7438], max: 0.0176[8674]
3:eval_score_p0 [50000]: avg: 0.1256 (± 0.0011), min: -0.7500[12930], max: 0.9167[8746]
3:eval_score_p1 [50000]: avg: -0.1256 (± 0.0011), min: -0.9167[8746], max: 0.7500[12930]
3:eval_score_p2 [50000]: avg: 0.1256 (± 0.0011), min: -0.7500[12930], max: 0.9167[8746]
3:eval_score_p3 [50000]: avg: -0.1256 (± 0.0011), min: -0.9167[8746], max: 0.7500[12930]
3:grad_norm [10000]: avg: 0.0264 (± 0.0000), min: 0.0141[7130], max: 0.0598[1167]
3:loss [10000]: avg: 0.0228 (± 0.0000), min: 0.0138[7438], max: 0.0355[8674]
3:reward [10000]: avg: 0.0000 (± 0.0000), min: -0.0107[2929], max: 0.0114[8696]
3:value_err [10000]: avg: 0.0115 (± 0.0000), min: 0.0069[3999], max: 0.0178[8674]
python main2.py num_thread=1 game=bridge env_actor_gen.params.gen_type=basic seed=1 \
trainer=selfplay actor_gen.params.batchsize=1 eval_only=true \
method=a2c agent.params.load_model=`pwd`/../bridge_models/agent-1610-0.63.pth \
baseline=console_eval game.params.feature_version=old \
game.params.train_dataset=`pwd`/../bridge_data/dda.db \
game.params.test_dataset=`pwd`/../bridge_data/test.db
Please check log of 3day model and 14day model competed against WBridge5. Here is an explanation of the log entry:
dealer is 0 [Vulnerability None] # The dealer is 0, no Vulnerability
[Deal "N:AK.J982.Q986.Q65 JT97.QT5.A5.AKJ7 Q82.AK743.32.T83 6543.6.KJT74.942"] # All 4 hands
parScore: -140
Seat ♠ ♥ ♦ ♣ HCP Actual Hand
0 2 4 4 3 12 ♠AK ♥J982 ♦Q986 ♣Q65
1 4 3 2 4 15 ♠JT97 ♥QT5 ♦A5 ♣AKJ7
2 3 5 2 3 9 ♠Q82 ♥AK743 ♦32 ♣T83
3 4 1 5 3 4 ♠6543 ♥6 ♦KJT74 ♣942
# At table 0, JPS at Seat 0 and Seat 2 (two AI doesn't know each other's hands), and WBridge5 at Seat 1 and Seat 3
# At table 1, JPS at Seat 1 and Seat 3, and WBridge5 at Seat 0 and Seat 2
# Bids in parentheses are from WBridge5 (the opponent)
Table 0, dealer: 0 1H (1N) P (P) P # Bidding. Seat 0 -> 1 -> 2 -> 3. JPS bids 1H, WBridge5 bids 1N (contract)
Table 1, dealer: 0 (1D) P (1H) P (2H) X (3H) P (P) P # Bidding. WBridge5 bids 1D, JPS bids P, Wbridge5 bids 1H, then 2H, then 3H (contract)
Table 0, Trick taken by declarer: 5, rawNSSeatScore: 100 # Result After DDS. 1N down 2, declarer (Wbridge5) loses 100 points and JPS won 100 points.
Table 1, Trick taken by declarer: 8, rawNSSeatScore: -50 # 3H down 1, declarer (WBridge5) loses 50 points and JPS won 50 points.
Final reward 0.166667 # Convert the two table scores into normalized IMPs (= IMP / 24)
To compute the overall performance given these logs, please run the following script:
./jps$ python compute_score.py --log_file ./logs/jps_3days.log
mean = 0.442, std = 0.1993448629244666
./jps$ python compute_score.py --log_file ./logs/jps_14days.log
mean = 0.628, std = 0.1944979317253668
Note that the original log had one bug that miscalculated the declarer (the declarer should be the first player calling for the strain of the final contract, rather than the last player who finalizes it). This affects the Final reward entry so a simple grep "Final reward" [log file] didn't give you the right answer. Instead we provide you with compute_score.py to compute the final score correctly, with the help of DDS table of the 1k games stored here.
In the folder ./dds is the double dummy utilities to quickly compute double dummy score given all hands and a contract.
In the folder ./vis there is visualization utility to visualize the bidding process given a complete bidding sequence and their action probabilities, as well as the DD table. ./vis/server.py is the server and ./vis/try.py is the client.
First initialize all submodules:
git submodule update --init --recursive
Then go to simple_game, and do the following to build
mkdir build
cd build
cmake ..
make
The executable jps is in ./build.
To start, in the build directory, run the following to get CFR1k+JPS solution for Mini-Hanabi. Log here:
./jps --game comm2 --iter 100 --iter_cfr 1000
You might run with --num_samples 1 to get the results for sampled-based version. E.g., run the following to get results of 100 trials:
for i in `seq 1 100`; do ./jps --game comm2 --iter 100 --iter_cfr 1000 --seed $i --num_samples 1; done > aa.txt
grep "CFRPure" aa.txt
Another example: Simple Bidding (N=16, d=3). Log here.
./jps --game simplebidding --seed 1 --iter 100 --N_minibridge 16 --iter_cfr 1000 --max_depth 3
There are a few tabular imperfect information collaborative games implemented:
comm: Simple Communication Game (Def. 1 in the paper)simplebidding: Simple Bidding (Def. 2 in the paper)2suitedbridge: 2-Suit Mini-Bridge (Def. 3 in the paper)comm2: Mini-Hanabi introduced in BAD paper.
Result
Results of sample-based approach
In build folder, do:
./jps --game=simplebidding --N_minibridge=4 --seed=2 > aa.txt
python ../load_strategies.py --log aa.txt
The output is:
Optimal policies
score: 2.1875
0 1 2 3
0 10 (0) 10 (1) 120 (2) 120 (2)
1 20 (0) 20 (2) 20 (2) 230 (4)
2 20 (2) 20 (2) 20 (2) 230 (4)
3 30 (0) 30 (4) 30 (4) 30 (4)
Note that "120" means P1 first bids 1, P2 then bids 2 and P1 bids 0 (Pass). The final contract is 2^{2-1} = 2, if card1 + card2 >= 2, then both of the players get reward 2 (shown in the parentheses), otherwise 0.
See the CONTRIBUTING file for how to help out.
JPS is under CC-BY-NC 4.0 license, as found in the LICENSE file.