valentinazangirolami/madrqn

Multi-Agent Deep Recurrent Q-Learning with Bayesian epsilon-greedy on AirSim simulator

Multi-Agent Deep Recurrent Q-Network

Description

A simple algorithm of Multi-Agent Deep Recurrent Q-Learning used to control speed and steer simultaneously using AirSim simulator. The structure of the process is illustrated on the following flowchart.

flowchart LR;
    A[Environment]-- Steering Reward -->C[Agent 1];
    A-->B{{Observation}}-->C;
    B--> D[Agent 2];
    A -- Speed Reward -->D;
    C-- steering angle --> G[Environment];
    D-- brake/throttle-->G;
    

The two agents take respectively actions without any connection, independently.

In future work, it would be interesting to evaluate a dependent structure between two agents to overcome the independent relation, which is less indicated in this task.

Prerequisites

• Python 3.7.6
• Tensorflow 2.5.0
• Tornado 4.5.3
• OpenCV 4.5.2.54
• OpenAI Gym 0.18.3
• Airsim 1.5.0

Related Work

• Deep Recurrent Q-Network