CNN-Based Game Agent for Chrome Dino Game Revisited

The automation of game-playing agents is a key domain for advancing AI algorithms. This research applies Convolutional Neural Networks (CNNs) to create an agent for the Chrome Dino game, overcoming the limitations of traditional rule-based systems. We address the core challenge of processing real-time visual data to execute precise jump and duck actions in a dynamic environment. Our solution is an end-toend CNN model that learns control policies directly from pixels, eliminating the need for hand-crafted features or internal game access. Experimental results confirm the model’s effectiveness, with performance metrics demonstrating proficient real-time decision-making and strong potential for visual-based automation in rapidly changing environments.