PhysicsGym: A Reinforcement learning interface for Partial differential equation control

This repository contains the final submission code for Master's thesis: Developing a Reinforcement learning interface for partial differential equation control.

Usage

Define the following methods specific to each problem

1. Initialisation

• obs_shape() : define the shape of your observation space
• action_shape() : define the shape of your action space
• physics : instantiate the physics object located in src/env/physics

2. Reset

• build_obs() : calculate the observation vector in this method.
• build_reward() : define reward calculation in this method.

3. Step

• action_transform() : define the transformation according to the problem, e.g. if you want to applying actions only at certain parts of the domain.

Experiments

• experiments/burgers_equation_experiments and experiments/heat_equation_experiments contains the PDE experiments for uncontrolled simulation, baseline agent, MPC agent and RL agent individually. The file ending in _eval compares all the three agents using multiple random initial states.