Developing Game AI for Classic Games Using Reinforcement Learning

New accomplishments have contributed widely to game development especially with respect to production of intelligent game playing entities. Most of the archetypal approaches for designing AI in games fail to work properly in complex and real-time environments with huge action space. This research addresses these challenges by utilizing reinforcement learning (RL) techniques to design AI models for two well-known games: Snake and Mario. The main challenge in the Snake game is to find the best action within a given state in a discrete state-action space. To solve this, we apply Q-learning, an uncomplicated, but highly effective RL algorithm that aims to accumulate the maximum reward by learning from previous experiences. The reason for Mario being more challenging than the other one is that Mario has a high dimensional input space and a continuous action space. For this, we use Proximal Policy Optimization (PPO), one of the most effective methods of reinforcement learning applicable in environments with the continuous decision space and stochastic nonlinear differential equations. This paper describes the design, deployment, and evaluation of these Reinforcement Learning agents where the major difficulties and the rectified methodologies are described. This study compares the efficiency of the application of Q-learning and PPO and shows how these approaches are more effective than the classical AI technologies in the context of gaming environments and bring valuable contributions toward further enhancement of game AI.