mimuw-rl / marl_team_based_testers Goto Github PK

We present Python scripts for testing agents in team-based MARL environments. Currently, the scripts are Unity ML-Agents compatible, however, it will work with any compiled team-based MARL environment with a similar Python interface. The scripts fill the gap for team-based Unity ML-Agents MARL, where a proper testing framework has been unavailable, and allows for comprehensive tests of agents in a team-based fashion, in environments like the cool Unity Dodgeball, and our new custom Hide & Seek and Predator-Prey environments.

Teams are defined through input .txt files that indicate the behavioral policy checkpoint for each agent. These checkpoints are in the Open Neural Network Exchange (ONNX) format, ensuring compactness and framework independence. Our script adeptly manages mixed discrete-continuous actions and varying agent numbers, efficiently handling in-game events like agent eliminations. We include example text teams and agent checkpoints in examples/. Notably, non-uniform team compositions can also be tested, allowing for testing out-of-distribution agent generalization capabilities.

The testing scripts will output a csv with results (the winning team identification, statistics like time hidden etc..) with a single row per match.

python unity_dodgeball_tester.py path_to_environment -e 1 -a 4 -t0 examples/dodgeball_elimination_team0.txt -t1 examples/dodgeball_elimination_team1.txt

where path_to_environment is the path to elimination.x86_64 -- the Dodgeball Elimination compiled test environment, downloadable from here, -t0 and -t1 are the paths to txt files with the team composition (t0=the blue team, t1=the purple team in this context), using the provided example checkpoints.

python unity_dodgeball_tester.py path_to_environment -e 1 -a 3 -t0 examples/dodgeball_ctf_team0.txt -t1 examples/dodgeball_ctf_team1.txt where path_to_environment is the path to ctf_1arena.x86_64 -- the compiled DB CTF test environment, downloadable from here, -t0 and -t1 are the paths to txt files with the team composition (t0=the blue team, t1=the purple team in this context), using the provided example checkpoints.

python3 hide_and_seek_tester.py path_to_environment -e 1 -a 3 -t0 examples/hide_seek_team0.txt -t1 examples/hide_seek_team1.txt where path_to_environment is the path to hidenseek.x86_64 -- the compiled Hide & Seek test environment, downloadable from here, -t0 and -t1 are the paths to txt files with the team composition (t0=the hiders team, t1=the seekers team in this context), using the provided example checkpoints.

python predator_prey_tester.py path_to_environment -e 1 -a 3 -t0 examples/predator_prey_team0.txt -t1 examples/predator_prey_team1.txt --game_params test_environments/predator_prey_configs/game_params.json --arena_params test_environments/predator_prey_configs/arena_params.json where path_to_environment is the path to predprey.x86_64 -- the compiled Predator-Prey test environment, downloadable from here, -t0 and -t1 are the paths to txt files with the team composition (t0=the prey team, t1=the predators team in this context), using the provided example checkpoints.

This environment requires additionally, the path to the config json files through the arguments --game_params, --arena_params, see for more details the Hide & Seek and Predator-Prey environments.

The tester results can be further processed to generate visualizations. Testers can be used to reproduce results presented in the submitted paper 'FCSP: Fictitious Co-Self-Play for Team-based, Multi-agent Reinforcement Learning', preprint available on request

Reach out in case of further inquiries.