Evolving Strategies For a Real-Time Strategy Game Using Genetic Algorithms
Translated title
Udvikling af Strategier til Real-Tids Strategi Spil Ved Brug af Genetiske Algoritmer
Authors
Pallesen, Steffan Bo ; Jensen, Mikkel Graarup ; Larsen, Nikolaj Dam
Term
4. term
Education
Publication year
2013
Submitted on
2013-06-07
Pages
88
Abstract
Spillere forventer stadig klogere modstandere i realtidsstrategispil for hver ny generation. Dette projekt undersøger, om genetiske algoritmer—søgemetoder inspireret af naturlig selektion—kan bruges til at evolvere mere udfordrende spil-AI. Det populære spil StarCraft bruges som testplatform, fordi det er veletableret i AI-forskning. To evolutionære tilgange blev overvejet til at udvikle en StarCraft-AI: en traditionel genetisk algoritme og en estimation-of-distribution-tilgang (der lærer en sandsynlighedsmodel over lovende løsninger). I dette arbejde blev kun den traditionelle genetiske algoritme implementeret og afprøvet. Resultaterne viser, at evolution kan frembringe en AI, der kan besejre en modstander, som hver gang spiller den samme faste (statiske) strategi. Disse resultater er et første skridt snarere end et færdigt produkt: yderligere forskning er nødvendig for at afgøre, om genetiske algoritmer kan udvikle komplette, tilpasningsdygtige AI’er egnet til kommercielle spil.
Players expect smarter opponents in real-time strategy (RTS) games with each new generation. This project asks whether genetic algorithms—search methods inspired by natural selection—can be used to evolve more challenging game AIs. We use the popular game StarCraft as a test bed because it is a well-established platform for AI research. Two evolutionary approaches were considered for building a StarCraft AI: a standard genetic algorithm and an Estimation of Distribution genetic approach (which learns a probability model of promising solutions). In this work only the standard genetic algorithm was implemented and evaluated. The results show that evolution can produce an AI that can beat an opponent that follows the same fixed (static) strategy in every match. These findings are a first step rather than a finished product: further research is needed to determine whether genetic algorithms can evolve complete, adaptable AIs suitable for commercial games.
[This abstract was generated with the help of AI]
Keywords
Documents