Optimizing an evolutionary approach to machine generated artificial intelligence for games
Authors
Constantin, Andrei Vlad ; Monastiridis, Konstantinos ; Cupit, Richard Alan
Term
4. term
Education
Publication year
2016
Submitted on
2016-05-24
Pages
94
Abstract
Dette speciale undersøger, hvordan man automatisk kan skabe og forbedre spil-AI ved hjælp af adfærdstræer (en udbredt måde at strukturere beslutningstagning i spil) og evolutionære algoritmer (metoder inspireret af naturlig selektion, der trinvis udvælger og forfiner bedre løsninger). Målet er en konceptdemonstration: et system, der kan frembringe alternative adfærdsmønstre i løbet af en enkelt gennemspilning, styret af evaluering af tidligere adfærd og med fokus på målbare forbedringer i præstation. Vi afprøver tilgangen i XCOM 2, et turbaseret taktikspil, ved at køre kampsimuleringer, hvor de genererede AI'er møder udviklernes standard-AI. Afslutningsvis viste et brugereksperiment, at den mest succesfulde maskingenererede AI vandt 50% af sine kampe mod standard-AI'en.
This thesis investigates how to automatically create and improve game-playing AIs using behavior trees (a common way to organize decision-making in games) and evolutionary algorithms (methods inspired by natural selection that iteratively select and refine better solutions). The goal is a proof of concept: a system that can generate alternative behaviors during a single playthrough, guided by evaluation of previous behaviors and aiming for measurable performance gains. We test the approach in XCOM 2, a turn-based tactics game, by running combat simulations where generated AIs face the developers' standard AI. Finally, in a user experiment, the most successful machine-generated AI won 50% of its matches against the standard AI.
[This abstract was generated with the help of AI]
Keywords
Documents