How tartrl/tikick works
TiKick is a multi-agent Google Research Football agent. Its documented evaluation workflow loads a trained actor, runs deterministic policy actions through parallel football environments, and saves replay data that can be rendered as video. Internally, the driver coordinates policy inference, while MAPPO, the policy network, and shared rollout storage provide learning components. The README also describes replay rendering; source excerpts do not show the full replay-writing path, so that relationship is documented rather than inferred. Several subsystem links rely on the SOURCE INDEX or README, not inspection of unsampled implementations.
Main entry points: football_evaluator.py Football evaluator · football_driver.py Football driver · base_evaluator.py Evaluation runner · football_env.py Football environment API
Components 21
Evaluation runtime
- Run configuration
tmarl/configs/config.py - Evaluation runner
tmarl/runners/base_evaluator.py - Football evaluator
tmarl/runners/football/football_evaluator.py - Football driver
tmarl/drivers/shared_distributed/football_driver.py - Vector environments
tmarl/envs/env_wrappers.py
Football simulation
- Multi-agent adapter
tmarl/envs/football/football.py - Football environment API
tmarl/envs/football/env/football_env.py - Football engine core
tmarl/envs/football/env/football_env_core.py - Scenario configuration
tmarl/envs/football/env/config.py - Scenario builder
tmarl/envs/football/env/scenario_builder.py - Football scenarios
tmarl/envs/football/scenarios
Policy and learning
- MAPPO algorithm
tmarl/algorithms/r_mappo_distributed/mappo_algorithm.py - MAPPO module
tmarl/algorithms/r_mappo_distributed/mappo_module.py - Policy network
tmarl/networks/policy_network.py - Action distributions
tmarl/networks/utils/act.py - Shared rollout buffer
tmarl/replay_buffers/normal/shared_buffer.py - Value normalization
tmarl/utils/valuenorm.py
Replay output
- Replay helpers
tmarl/envs/football/env/script_helpers.py - Replay video renderer
scripts/football/replay2video.py - Saved replay
Other components
- Evaluator
Connections 25
- Evaluator → Football evaluator: starts evaluation
tmarl/runners/football/football_evaluator.py - Football evaluator → Evaluation runner: extends
tmarl/runners/football/football_evaluator.py - Evaluation runner → Run configuration: loads options
tmarl/runners/base_evaluator.py - Evaluation runner → Football driver: initializes driver
tmarl/runners/football/football_evaluator.py - Football evaluator → Multi-agent adapter: selects environment
tmarl/runners/football/football_evaluator.py - Football evaluator → Vector environments: creates vector envs
tmarl/runners/football/football_evaluator.py - Football driver → MAPPO algorithm: builds trainer
- Football driver → Shared rollout buffer: uses rollout storage
- Football driver → MAPPO module: requests actions
tmarl/drivers/shared_distributed/football_driver.py - Football driver → Vector environments: steps environments
tmarl/drivers/shared_distributed/football_driver.py - Vector environments → Multi-agent adapter: steps wrapped env
- Multi-agent adapter → Football environment API: creates environment
tmarl/envs/football/football.py - Football environment API → Football engine core: uses engine core
tmarl/envs/football/env/football_env.py - Football environment API → Scenario configuration: uses configuration
tmarl/envs/football/env/football_env.py - Scenario configuration → Scenario builder: uses builder
tmarl/envs/football/env/config.py - Scenario builder → Football scenarios: loads scenarios
- MAPPO algorithm → MAPPO module: updates module
- MAPPO module → Policy network: constructs policy
tmarl/algorithms/r_mappo_distributed/mappo_module.py - Policy network → Action distributions: uses action layer
tmarl/networks/policy_network.py - MAPPO algorithm → Value normalization: uses normalization
tmarl/algorithms/r_mappo_distributed/mappo_algorithm.py - Shared rollout buffer → Value normalization: denormalizes returns
tmarl/replay_buffers/normal/shared_buffer.py - Football evaluator → Saved replay: saves replay
README.md - Replay video renderer → Replay helpers: uses helpers
scripts/football/replay2video.py - Evaluator → Replay video renderer: renders replay
README.md - Replay video renderer → Saved replay: reads replay
scripts/football/replay2video.py