MADDPG_simpletag

MADDPG environment to solve openai's 'simple_tag' environment.
Three(default) predators chase a preyer for reward(10) plus shaped reward(distance of predators and preyers). Three predator choose action with MADDPG algorithms and the preyer acts with uniform distribution from -1. to 1.

Dependency

• pytorch==1.0.1
• tensorboardX
• Use my environment on envs or...
  • Install the OpenAI's environment and edit some codes

# environment.py L29

# self.discrete_action_space = True
self.discrete_action_space = False

# simple_tag.py L92
def agent_reward(self, agent, world):
    # Agents are negatively rewarded if caught by adversaries
    rew = 0
    # shape = False
    shape = True

# simple_tag.py L118
def adversary_reward(self, agent, world):
    # Adversaries are rewarded for collisions with agents
    rew = 0
    # shape = False
    shape = True

Getting started

Train

python train.py --tensorboard

Result

Acknowledgement

• shariqiqbal2810's Pytorch implementation (Most motivated)
• Kostrikov's Pytorch implementation