#Maze Task - Reinforcement Learning
This repo contains my reinforcement learning experiment on the maze task domain.
A reinforcement learning agent is learned to reach a given goal position in a maze.
Tabular Q-learning is used for learning the policy.
A maze of size nXn, with one goal position, starting from any random position in the maze, an agent has to reach to the goal position.
See the image given below -- grey cells are the feasible states to the agent, white cell is the goal position. Action allowed in the feasible states are - up, right, bottom, left with action should lead to an another feasible state.
Dependencies:
- python3
- numpy
- matplotlib
Run from the root of the directory:
python main.py
It will plot the final policy in the policies folder. The policies folder contains policies after every 50 episoded, it can be obtained by turning variable FLAG_policy in main.py True.
The gif of policy leaned is:
