Development and Performance Analysis of an AI based Agent to Play Computer Games using Reinforcement Learning Techniques

Exploring the potential of Artificial Intelligence (AI) in understanding and improving the tactics to play computer games is one of the most trending subjects being explored in recent times. Whenever a new game hits the market, users might struggle to play the game, but creating an AI agent would help understand how the game must be played. The purpose of developing AI agents to play video games are Self- improvement, Learning and Flexibility. Reinforcement learning (RL) has emerged as a promising approach for tackling intricate tasks across different fields, including the realm of gaming. Among the multitude of games, Flappy Bird has gained popularity due to its challenging nature, proving to be a formidable environment for traditional artificial intelligence techniques. This project aimed to develop an AI agent capable of playing Flappy Bird by employing three distinct RL algorithms: Deep Q-learning Network (DQN), State-Action-Reward-State-Action (SARSA), and Double Deep Q-learning Network (DDQN).Performance of each algorithm is evaluated by measuring the rewards and scores achieved in the game. Results show that DDQN outperformed DQN and SARSA in terms of achieving higher scores and rewards. An epsilon-greedy strategy is incorporated into the training process to effectively balance the exploration and exploitation aspects. Additionally, the utilization of experience replay has significantly enhanced the learning efficiency of the reinforcement learning algorithms by storing and reusing previous experiences. Analysis of the proposed system demonstrates the potential of RL algorithms in solving challenging tasks in the gaming domain. Future work could focus in investigating how to transfer learning from one environment to another and how to optimize hyperparameters for each algorithm.