Yoshi-0921 / EXP

Minimal frameworks for multi-agent reinforcement learning with deep neaural network.

EXP-MADRL

0. What is this repo for?

This repository is for experiments to investigate unknown features and potential advantages of multi-agent reinforcement learning with deep neaural network. Code is basically written in Python using PyTorch .

GIF cited from here

1. Environments

Simple empirical environment has been implemented so far.

1.0 Maps

Setting up such appropriate environment is quit important. Agents interact with the environment and the other agents in the following grid map which has several rooms of 9x9, separated by 2-width hallways and 1-width wall.

1.1 Exp1, Exp2

Agents and events are randomly put inside the grid map. Agents aim to approach the events as fast as possible.

Action	State	Reward
0: Stay	obs1: [agent.state.p_pos]	rew1 : -min(dist) / (SIZE_X * num_agents)
1: Right	obs2: entity_pos	rew2 : -(1 / num_agents) if is_collision
2: Up	obs3: other_pos	rew1 + rew2
3: Left	np.concatenate(ob1 + obs2 + obs3)	-
4: Down	-	-

1.2 Exp3

Still in progress.

2. Deep Reinforcement Learning

2.1 Deep Q-Network

2.2 Deep Deterministic Policy Gradient