Game-Theory-Final-Project

Outlook

algos.py

Contains the implementation of all the RL algorithm. Each algorithm has two version :

• One version using Tables for the value-functions
• One version using neural networks to approximate the values of the value-functions

agent.py

Contains the implementation of the agent used in the 5 experiments

experimentConf.py

Contains all the parameters that we used for the algorithms/ensemble methods for each experiment

mazes.py

Contains the implementations off all the environments (mazes) needed for the 5 experiments as well as functions to generate such environments.

clusterRunner.py

Script used to run the experiments on a cluster and distribute the trials across multiple cores.

Tasks

Neural nets

• Q-learning
• SARSA
• Actor-Critic
• QV-learning
• ACLA

Belief State

• Belief State
• Maze observations

Experiments

• Exp 1 (Simple maze + base algo)
• Exp 2 (Partially obsebable maze + neural net)
• Exp 3 (Dynamic obstacles maze + neural net)
• Exp 4 (Dynamic Goal maze + neural net)
• Exp 5 (Generalized maze + neural net)

Base Algorithms

• Q-learning
• SARSA
• Actor-Critic
• QV-learning
• ACLA

Ensemble methods

• Majority voting
• Rank voting
• Boltzmann multiplication
• Boltzmann addition

Environments

• Simple Dyna maze (9x6)
• Dyna maze with Dynamic Goal (9x6)
• Dyna maze with dynamic obstacles (9x6)
• Generalized maze (9x6)