'Combinatory Tests of AI Techniques for Common Tasks in Computer Games'
Authors
Coltau, Pelle ; Jacobsen, Jens Juul ; Jensen, Brian
Term
4. term
Education
Publication year
2006
Abstract
I dette projekt undersøger vi, hvordan læringsbaserede AI-metoder kan bruges i computerspil, ved at bygge et modulært system, der spiller brætspillet Risk. Systemet opdeler almindelige spilopgaver, som at forstå spillets tilstand og planlægge træk, så vi kan afprøve forskellige metoder i hver del. Vi evaluerer scripting (regelbaserede, håndskrevne regler), beslutningstræer, neurale netværk, bayesianske netværk og naive Bayes-klassifikatorer. En scriptet Risk-spiller fungerer som lærer og skaber træningsdata til læringsmetoderne. Vi beskriver, hvordan vi trænede hver metode, og de praktiske udfordringer vi mødte. Vi foreslår også en måde at måle, hvor vigtigt hvert modul er, samt en testplan, der effektivt sammenligner mange AI-spillere uden at køre alle mulige opgør. Den bedste spiller er en hybrid, der kombinerer forskellige metoder, hver brugt der, hvor den klarer sig bedst. I flere tilfælde overgår de lærte komponenter endda deres scriptede lærer. Til sidst drøfter vi, hvilke spilopgaver de enkelte metoder sandsynligvis egner sig bedst til.
This project tests how learning-based AI methods can be applied to computer games by building a modular system that plays the board game Risk. The system separates common game tasks—such as understanding the game state and planning moves—so we can try different methods on each part. We evaluate scripting (hand-written rules), decision trees, neural networks, Bayesian networks, and naive Bayes classifiers. A scripted Risk player serves as a teacher that generates training data for the learning methods. We describe how we trained each method and the practical issues we encountered. We also propose a way to measure how important each module is and a test plan that compares many AI players efficiently, without running every possible matchup. The best overall player is a hybrid that combines different methods, each used where it performs best. In several cases, the learned components even outperform their scripted teacher. We conclude by discussing which kinds of game tasks each method is likely to suit best.
[This abstract was generated with the help of AI]
Documents