Multi_Agent_Soft_Actor_Critic

A Pytorch Implementation of Multi Agent Soft Actor Critic

Project Details

• The environment consists of multiple agents where the task of the agent hit the ball and keep it in the air without allowing it to fall on the ground.
• The current state of the environment is represented by 24 dimensional feature vector which conist the position of the ball and speed of the ball
Action space is continous and thus it represent by a vector with 2 numbers, corresponding to position of the bat ranging between -1 and 1 in each dimension.
• A reward of +0.1 is provided for time the agent's hits the ball and -0.1 if the agent miss it or shoots the ball away from the court.
• The task is episoidic, and in order to solve the environment, the agent must get an average score of +0.5 over 100 consecutive episodes

Technical Dependencies

1. Python 3.6 :
2. PyTorch (0.4,CUDA 9.0) : pip3 install torch torchvision
3. ML-agents (0.4) : Refer to ml-agents for installation
4. Numpy (1.14.5) : pip3 install numpy
5. Matplotlib (3.0.2) : pip3 install matplotlib
6. Jupyter notebook : pip3 install jupyter
7. Download the environment from here and place it in the same folder as that of Tennis.ipynb file

Network details

• Value - function Network
• Entropy Regularization
• Two Action - Value Networks
• Centralized Training
• Decentralized Execution

Installation Instructions :

step 1 : Install all the dependencies
step 2 : git clone https://github.com/adithya-subramanian/Multi_Agent_Soft_Actor_Critic.git
step 3 : jupyter notebook
step 4 : Run all cells in the Tennis.ipynb file

Acknowledgment

Certain parts of SAC.py,model.py and Tennis.ipynb has been partially taken from the Udacity's deep reinforcement learning Nanodegree.