Human Guided Ensemble Learning in StarCraft

Timothy Verstraeten, Roxana-Teodora Radulescu, Yannick Jadoul, Tom Jaspers, Robrecht Conjaerts, Tim Brys, Anna Harutyunyan, Peter Vrancx, Ann Nowe

Research output: Chapter in Book/Report/Conference proceedingConference paper

Abstract

In reinforcement learning, agents are typically only rewarded based on the task requirements. However, in complex environments, such reward schemes are not informative enough to efficiently learn the optimal strategy. Previous literature shows that feedback from multiple humans could be an effective and robust approach to guide the agent towards its goal. However, this feedback is often too complex to specify beforehand and is generally given during the learning process. We introduce real-time human guided ensemble learning, in which feedback from multiple advisers is learned simultaneously with the agent's behaviour. We evaluate our approach in a small scale one-on-one combat scenario in StarCraft: Brood War. Our results show that a single expert adviser can provide proper guidance, while using groups of multiple advisers does not improve the convergence speed. In future work, we will investigate an alternative to the tile-coding approximator in order to effectively incorporate advice from multiple humans.
Original languageEnglish
Title of host publicationProceedings of the 16th Adaptive Learning Agents Workshop (ALA) at AAMAS 2016
Pages99-105
Number of pages7
Publication statusPublished - 9 May 2016
Event16th Adaptive and Learning Agents Workshop, ALA 2016 - , Singapore
Duration: 9 May 201610 May 2016

Conference

Conference16th Adaptive and Learning Agents Workshop, ALA 2016
Country/TerritorySingapore
Period9/05/1610/05/16

Fingerprint

Dive into the research topics of 'Human Guided Ensemble Learning in StarCraft'. Together they form a unique fingerprint.

Cite this