RL_maze

Escape a maze using (deep) reinforcement learning

Generates a grid maze, with start, finish and walls, which an agent has to navigate. The agent can move through the maze by performing actions and learning from them.

How-to

Pre-requisites

• Install poetry for dependency management and install venv: poetry install
• Open poetry shell: poetry shell Then run the desired Python commands.

Configuration

A complete configuration should be in root/config/ and consists of:

• Config file: Maze settings, parameters, training
• Maze file: Text file containing a maze with walls (1) and hallways (0) and other features.

Choose settings in root/config/<config_file>:

• epochs
• learning rate $\alpha$
• discount rate $\gamma$
• exploration vs. exploitation $\epsilon$
• rewards

Custom mazes and configurations can easily be added.

Run

RLMaze can easily be used as follows:

from rlmaze.rlmaze import RLMaze
RLMaze(<config_file>).escape_maze()

to train on a maze. The maze can be visualized with:

Utils().visualize_maze(<config_file>)

Some test cases can be found in `root/tests/' and run with:

python tests/test_maze_diagonal.py

or

python tests/test_maze_13x13.py

Figures

Diagonal maze

Fig 1: Diagonal maze layout with start (S) and finish (F).

Fig 2: Diagonal maze training (deep RL). Convergence to optimal number of steps (8).

Fig 2: Diagonal maze training (Classic Q-learning). Convergence to optimal number of steps (8).

13x13 maze

Fig 3: Diagonal maze layout with start (S) and finish (F).

Fig 4: 13x13 maze training (Classic Q-learning). Convergence to optimal number of steps (20).