Half Field Offense in Robocup 2D Soccer with reinforcement learning

Framework

Algorithm

Independent PA-DDPG, i.e., there is no centralized critic or communication. Each agent acts asynchronously. Multi-agent scenarios can be implemented using redis.

Reward function

+1 if goal, else 0 (sparse reward)

Offense action space

Similar to MAPQN, there are 3 mid-level parameterized actions (kick to, move to, dribble to) and a discrete high-level action (shoot) for offense players to choose.

Observation space

Low level features in HFO.

Modules

Results

• Evaluation frequency: every 500 episodes
• Evaluation length: 1k episodes

1v1 against the world champion HELIOS

2v1 against HELIOS

2v2 against HELIOS

Examples

Use command --help for more parameter settings.

Connect to hfo-server

You can adjust the number of players in training and evaluation accordingly (e.g. for self-play).

python connect.py --offense-agents 2 --defense-agents 0 --defense-npcs 1 --server-port 6000
redis-server

Training

Start a learner for an agent:

python learner.py --tensorboard-dir agent1 --save-dir agent1

Evaluation

Start an evaluator for an agent:

python evaluator.py --tensorboard-dir agent1 --save-dir agent1 --episodes 20000

2v2 example

PDDPG 2v2 training:

python connect.py --offense-agents 2 --defense-agents 0 --defense-npcs 2 --server-port 6000
python learner.py --tensorboard-dir agent1 --save-dir agent1
python learner.py --tensorboard-dir agent2 --save-dir agent2

Evaluate PDDPG 2v2 model:

python connect.py --offense-agents 2 --defense-agents 0 --defense-npcs 2 --server-port 6000
python evaluator.py --tensorboard-dir agent1 --save-dir agent1
python evaluator.py --tensorboard-dir agent2 --save-dir agent2

Citing

If this repo helps you, please star it.

Reference

The code in this repo has refered to HFO, MP-DQN, PA-DDPG, gym-soccer.