Using Poisson Markov Models to Predict Game States in StarCraft
Author
Hesselager-Olesen, Anders
Term
4. term
Education
Publication year
2014
Submitted on
2014-03-06
Pages
76
Abstract
Denne opgave undersøger, om det er muligt at forbedre genkendelsen af strategier og forudsigelsen af fremtidige handlinger i realtidsstrategispillet StarCraft ved at kombinere forskellige observationstyper i skjulte Markov-modeller. En skjult Markov-model er en statistisk model, der udleder skjulte tilstande (fx en spillers strategi) ud fra observerede data over tid. Vi bruger blandede observationer: kategoriske hændelser modelleres med en multinomialfordeling, og optællinger over tid modelleres med en Poisson-fordeling. Vi præsenterer teorien for begge fordelinger og for deres kombination i én model. Data fra StarCraft analyseres for at vurdere, om en Poisson-fordeling er passende til at beskrive produktionen af kampenheder. Til sidst gennemfører vi eksperimenter for at måle modellens forudsigelsesnøjagtighed og for at evaluere den mest sandsynlige rute gennem modellens tilstandsrum.
This project investigates whether combining different observation types in Hidden Markov Models can improve recognizing strategies and predicting future actions in the real-time strategy game StarCraft. A Hidden Markov Model is a statistical model that infers unobserved states (such as a player's strategy) from observed data over time. We use mixed observations: categorical events are modeled with a multinomial distribution, and counts over time are modeled with a Poisson distribution. We present the theory behind each distribution and how to combine them in a single model. StarCraft data are analyzed to assess whether a Poisson distribution is appropriate for modeling the production of combat units. Finally, we run experiments to measure the model’s predictive accuracy and to evaluate the most likely path through the model’s state space.
[This abstract was generated with the help of AI]
Keywords
Documents