Online Build-Order Optimization for Real-Time Strategy Agents using Multi-Objective Evolutionary Algorithms

Abstract

The investigation introduces a novel approach for online build-order optimization in real-time strategy (RTS) games. The goal of our research is to develop an arti cial intelligence (AI) RTS planning agent for military critical decision- making education with the ability to perform at an expert human level, as well as to assess a players critical decision- making ability or skill-level. Build-order optimization is modeled as a multi-objective problem (MOP), and solutions are generated utilizing a multi-objective evolutionary algorithm (MOEA) that provides a set of good build-orders to a RTS planning agent. We de ne three research objectives: (1) Design, implement and validate a capability to determine the skill-level of a RTS player. (2) Design, implement and validate a strategic planning tool that produces near expert level build-orders which are an ordered sequence of actions a player can issue to achieve a goal, and (3) Integrate the strategic planning tool into our existing RTS agent framework and an RTS game engine. The skill-level metric we selected provides an original and needed method of evaluating a RTS players skill-level during game play. This metric is a high-level description of how quickly a player executes a strategy versus known players executing the same strategy. Our strategic planning tool combines a game simulator and an MOEA to produce a set of diverse and good build-orders for an RTS agent. Through the integration of case-base reasoning (CBR), planning goals are derived and expert build- orders are injected into a MOEA population. The MOEA then produces a diverse and approximate Pareto front that is integrated into our AI RTS agent framework. Thus, the planning tool provides an innovative online approach for strategic planning in RTS games. Experimentation via the Spring Engine Balanced Annihilation game reveals that the strategic planner is able to discover build-orders that are better than an expert scripted agent and thus ac

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Mar 27, 2014
Accession Number
ADA599118

Entities

People

  • Jason M. Blackford

Organizations

  • Air Force Institute of Technology

Tags

Communities of Interest

  • Energy and Power Technologies
  • Human Systems

DTIC Thesaurus Topics

  • Air Force
  • Algorithms
  • Artificial Intelligence
  • Bayesian Networks
  • Computational Science
  • Computer Programming
  • Data Mining
  • Evolutionary Algorithms
  • Genetic Algorithms
  • Information Science
  • Machine Learning
  • Mathematical Models
  • Military Education
  • Operating Systems
  • Students
  • Supervised Machine Learning
  • Warfare

Readers

  • Distributed Systems and Data Platform Development
  • Game Theory.
  • Team-Based Human-Centered Cognitive Task Decision Making and Information Performance.