Multi-Agent Reinforcement Learning for Fast Timescale Demand Response of Flexible Loads

This is the original version of the code that was used to produce the results of the Multi-Agent Reinforcement Learning for Fast-Timescale Demand Response of Residential Loads paper.

The officially maintained version of the code is refactored and better documented, and available here: https://github.com/ALLabMTL/marl-demandresponse

Dependencies:

• ray-rllib
• numpy
• pandas
• wandb
• matplotlib
• pytorch
• noise
• sympy
• pyglet
• cvxpy (only needed by the MPC agent - you can disable it if necessary)

To run the code

You need to have logged in to wandb before (otherwise, you can simply remove all wandb calls from the code).

To train an agent:

python main.py --exp [EXP NAME]

See the cli_train() definition in cli.py to know the other arguments. See config.py to know the default arguments.

To deploy:

python main-deploy.py --agent [AGENT NAME] --agent_type [AGENT TYPE]

See the cli_deploy() definition in cli.py to know the other arguments. See config.py to know the default arguments.