AI in Games
Agent Enhancement and RL Algorithms
About The Module
In AI in Games module, I have worked and learned about various aspects of Machine Learning, Deep learning and Reinforcement Learning. In the first half of this module, I have acquired a knowledge about the various types of AI agents used in games namely Monte Carlo Tree Search (MCTS), Rolling Hill Evolutionary Algorithm (RHEA), One-Step LookAhead (OSLA), & various searches such as MinMax, Alpha Beta etc. Also, learned about the different pruning techniques for tree search such as First Play Urgency, Unprunning/Widening tree search, Portfolio Search etc.
In the other half of the module, I have acquired a hands-on knowledge about fundamentals of Reinforcement Learning and Deep Reinforcement Learning. How RL algorithms aims to find the optimal policy using the Markov Decision Process, various algorithms used to solve the Markov Decision Processes such as Value Iteration, Policy Iteration Policy Evaluation, In-Place Policy Evaluation etc., learned about the difference between Tabular Model-Free algorithms and Non-Tabular Model-Free algorithms, how different types of algorithm works such as Sarsa algorithm, Q-learning algorithm. Also, learned about the Deep Reinforcement Learning algorithms and their application in real world scenarios.
Agent Enhancement
For the first half of the module's coursework, my other two mates & I worked on the enhancement of an MCTS AI agent, experimenting the agents with different settings for the Sushi Go game, & writing a research paper on the obtained results of the experimentation. Through the course of this coursework, I have gained a knowledge about how different agents work in games and how we can improve their performance. Furthermore, I have written a heuristic, with my other mate, of the game Sushi Go which improves the performance of the MCTS agent in the game.
Link to the Paper: Agent Enhancement Experiment Paper
Reinforcement Learning Algorithms
For the other half of the module's coursework, I have learned about the RL algorithms and written a code of the various algorithms, for the Frozen Lake Game using python, such as Value Iteration, Policy Iteration, Sarsa control, Q-learning control, Sarsa control with linear function approximation, Q-learning control with linear function approximation and deep Q-learning learning.
Note: Due to Code of Honor, I'm not entitled to show the code outside the University premises.
