Machine learning state evaluation in prismata

Rory Campbell
Strategy games are a unique and interesting testbed for AI protocols due their complex rules and large state and action spaces. Recent work in game AI has shown that strong, robust AI agents can be created by combining existing techniques of deep learning and heuristic search. Heuristic search techniques typically make use of an evaluation function to judge the value of a game state, however these functions have historically been hand-coded by game experts. Recent results have shown that it is
more » ... ossible to use modern deep learning techniques to learn these evaluation functions, bypassing the need for expert knowledge. In this thesis, we explore the implementation of this idea in Prismata, an online strategy game by Lunarch Studios. By generating game trace training data with existing state-of-the-art AI agents, we are able to use a Machine Learning (ML) approach to learn a new evaluation function. We trained several evaluation models with various parameters in order to compare prediction time with prediction accuracy. To evaluate the strength of our learned model, we ran a tournament between AI players which differ only in their state evaluation strategy. The results of this tournament demonstrate that our learned model when combined with the existing Prismata Hierarchical Portfolio Search system, produces a new AI agent which is able to defeat the previously strongest agents. A subset of the research presented in this thesis was the subject of a publication in the Artificial Intelligence and Interactive Digital Entertainment (AIIDE) 2019 Strategy Games Workshop [1].
doi:10.48336/2age-wf61 fatcat:gpxjuirmyjerfkvhz7ejplsdai