Reinforcement Learning Project

This project simulates user interactions to predict posture changes based on window position changes on the screen using reinforcement learning (RL) algorithms.

Environment Setup

ScreenEnv Class

  • Grid Representation: A 6x6 grid (36 positions) representing the screen.
  • Window Position: The window starts at a specific position and can move within the grid.
  • State and Action Spaces: Defines the possible states for the window and actions for the user.

Agents

Expected SARSA Agent

  • Purpose: Effective in stochastic environments.
  • Key Features: Balances exploration and exploitation using an epsilon-greedy strategy and updates Q-values based on expected future rewards.

Double Q-Learning Agent

  • Purpose: Reduces overestimation bias in stochastic environments.
  • Key Features: Maintains two Q-tables to decouple action selection from evaluation, providing more stable learning.