Optimization of behavioral parameters of Artificial Intelligence agents created by UtilityAI tool
Author
Caniga, Vladimir
Term
4. term
Education
Publication year
2019
Pages
35
Abstract
Dette projekt har til formål at gøre arbejdsgangen mere smidig ved design af en AI‑agent bygget med UtilityAI fra Apex Game Tools. Sådanne agenter afhænger af mange numeriske parametre, som skal finjusteres, for at agenten opfører sig som ønsket. Vi foreslår at definere en belønningsfunktion—en regel, der giver højere point for ønsket adfærd—og derefter bruge en optimizer til at justere parametrene, så belønningen maksimeres. Vi anvender Bayesiansk optimering, en statistisk metode der effektivt finder gode parameterværdier med relativt få forsøg. Tilgangen evalueres på spillet Survival Shooter og på to kendte, kunstigt konstruerede funktioner.
This project aims to streamline the workflow for designing an AI agent built with UtilityAI by Apex Game Tools. Such agents rely on many numeric parameters that must be carefully tuned so the agent behaves as intended. We propose defining a reward function—a rule that assigns higher scores to desirable behavior—and then using an optimizer to adjust the parameters to maximize that reward. We adopt Bayesian optimization, a statistical method that efficiently searches for good parameter settings with relatively few trial runs. The approach is evaluated on the Survival Shooter game and on two known, artificially constructed functions.
[This abstract was generated with the help of AI]
Keywords
Documents