Co-evolving Distributed Control for Heterogeneous Agents in RTS Games

Loading...
Thumbnail Image

Authors

Adhikari, Navin K.

Issue Date

2019

Type

Thesis

Language

Keywords

Artificial Intelligence , Coevolution , Distributed Control , Evolutionary Algorithms , Potential Fields , RTS Micro

Research Projects

Organizational Units

Journal Issue

Alternative Title

Abstract

We investigate competitive and co-operative co-evolutionary approaches to generating transparent distributed control for teams of heterogeneous agents in RTS games. RTS games provide a challenging test-bed for AI researchers as they simulate many fundamental AI research problems in a virtual world. In this thesis, we use an open-source RTS game engine called FastEcslent and the popular RTS game, Starcraft II as our test-beds. We represent the problem of generating transparent distributed control in RTS games as a set of numerical parameters that co-evolve to optimize distributed control for agents in multiple skirmish scenarios. This thesis makes three contributions to research in distributed transparent control.First, we remove the need for having a high performing opponent to evolve against by using competitive co-evolution to generate different individual control behaviors for agents working towards single-objective or multi-objective goals from scratch. We then use case-injection to transfer competitively co-evolved behavior from FastEcslent, which runs fast and enables many more fitness evaluations in a reasonable time, to Starcraft II.Second, we remove the need for having one general representation for evolving distributed control for heterogeneous agents by co-operatively co-evolving individual groups of similar agents in Starcraft II. Each group has a different evolutionary representation but shares a common fitness evaluation. Third, we also investigate a new representation for generating such control for ranged agents and show improvement over our previous representation. Results show that we can co-evolve winning distributed control in skirmishes across two different RTS games and with multiple representations. Results also show that we can competitively co-evolve higher performance, faster, with case-injection. We believe these results indicate the viability of our co-evolutionary approaches and representations for generating high quality, transparent, distributed control for heterogeneous agents in RTS games and that case-injection can lead to skill transfer across similar environments.

Description

Citation

Publisher

License

Journal

Volume

Issue

PubMed ID

DOI

ISSN

EISSN