|Imitation Learning: Building Practical Agents to Test and Explore a First-Person Shooter
|Track / Format:
Did you know free users get access to 30% of content from the last 2 years?
|The talk introduces style-centric autoplay agents that game developers can train quickly and on a budget to facilitate testing and evaluating the gameplay of a title under development. The focus on the practical and stylistic aspects suggests a simple approach based on the Markov model. Markov agents are trainable interactively and can incorporate new game features without re-training the entire model. While such agents efficiently capture the demonstrated gameplay style, they may require fallback heuristics to address possible lack of performance, or the game states not explicitly present in organic play-throughs. Imitation learning on the data bootstrapped from the enhanced with heuristics Markov agent allows training of a compact computationally efficient DNN model suitable for automated game evaluation and testing. A generic First-Person Shooter game example provides a practical context for the presentation. The GitHub repository illustrates Markov agents in more detail and uses OpenAI gym environments for an interactive demo.