Maze Reinforcement

Implementation of Reinforcement Learning using Reinforcement.js in my Maze Generation Algorithm.

Environment

Main api for the environment is taken from Maze Generation Algorithm. But, in the world (implemented for reinforcement learning) has following things:

1. reset to reset the game state.
2. reward return reward from current state, action and next state.
3. sampleNextState calculate next state and reward from an action suggested by agent.
4. allowedActions return allowed actions at current cell.

Beside everything in world call are helpers.

Reward

The rewards given to the agent are under following conditions:

1. -0.01 on every move. So that agent don't stuck on same place.
2. When agent solves the puzzle give positive reward equal to the size of puzzle.
3. -0.1 on start, top right and bottom left tiles.

Specs

Spec	Value
Discount Factor (gamma)	0.9
Epsilon-greedy Policy	0.2
Learning Rate (alpha)	0.3
Eligibility trace decay (lambda)	0
Replacing Traces	true
Number of planning steps per iteration	50
Smooth Policy Update	true
Learning Rate for Smooth Policy (beta)	0.1