Training an agent using Reinforcement Learning to crack a labyrinth
We analyze the SARSA and Q-learning algorithms applied on a 10x10 grid where an agent is moving trying to reach the goal. In particular different parameters are tuned for both algorithms (episode-length,lambda,alpha,epsilon) searching for the best combinations and providing a comparison between SARSA and Q-learning. Finally different obstacles are fed into the environment in order to better analyze the agent behavior.
Walls are placed in the grid which gave a negative reward of -0.1 and they cannot be passed. Sand damage the robot circuits giving him a reward of -0.3, while going outside of the grid results in a negative reward of -0.6.
The agents can take 5 different actions (Up,Down,Left,Right,Stay).
For the following tests a Grid Search has been implemented in order to find the better configurations of parameters. (alpha discount factor, epsilon, episode length).
In the training.pyyou can enable the best parameters search with the paramsearch mode, or you can run in the agent mode.
Moreover you can tune the following parameters
episodes --> # number of training episodes
x --> #horizontal size of the box
y --> #vertical size of the box
goal --> #objective point
softmax --> #true use Softmax policy
sarsa --> #true to use Sarsa instead of Q-Learning
discount --> #discount factor applied
epsilon --> #epsilon greedy policy Labyrinth visual example and the path selected by the trained agent.
