290 likes | 430 Views
Generating Diverse Opponents with Multi-Objective Evolution. Alexandros Agapitos, Julian Togelius , Simon M. Lucas, J¨urgen Schmidhuber and Andreas Konstantinidis Presented by Patoka Amir. Overview. Introduction Objectives Multi-objective evolutionary algorithms Results Future work.
E N D
Generating Diverse Opponents with Multi-Objective Evolution Alexandros Agapitos, Julian Togelius, Simon M. Lucas, J¨urgenSchmidhuber and Andreas Konstantinidis Presented by Patoka Amir
Overview • Introduction • Objectives • Multi-objective evolutionary algorithms • Results • Future work
Introduction • Easy construction of game AI (F.E.A.R.winer of GameSpot Best Artificial Intelligence) • Industry shifts focus to building interesting, divers and believable CI.
Introduction (Cont) • DIFFICULTY LEVELS: • Barbarians • Free Units • Research • Maintenance Costs • Health and Happiness • Artificial Intelligence Penalties • AI Freebies • Tribal Villages
Introduction (Cont) • Predictable AI • When Priorities Go Wrong!: NPC investigates a burning barrel that was thrown by the player and landed nearby. The barrel subsequently explodes while the NPC is nearby looking at it.
Introduction (Cont) • Propose a general approach to creating diverse and interesting NPC behaviors using Multi-objective evolutionary algorithms (MOEA) in combination with a number of partly conflicting behavioral fitness measures.
Overview • Introduction • Objectives • Multi-objective evolutionary algorithms • Results • Future work
Objectives • Optimize a genetically programmed car controller to exhibit: • Aggressiveness. • Opponent weakness exploitation.
Objectives (Cont.) • Environment: • A 2D simulator, modeling a radio controlled toy car (three possible drive and steering modes). • A track consisting of walls, a chain of waypoints and a set of staring points and directions (subject to random alteration). • A reasonable model of car dynamics, collisions. • A competitor (with an incrementally evolved general controller).
Objectives (Cont.) • Controller employ two expression trees representation (driving and steering) containing: • Standard arithmetic and trigonometric functions. • Formal parameters representing car state as viewed by first person sensors.
Objectives (Cont.) • Behavioral fitness measures: • Absolute progress. • Relative progress. • Maximum speed. • Progress variance. • # Steering changes. • # Driving changes. • Wall collisions. • Competitor proximity. • Max Car collisions. • Min Car Collisions.
Objectives (Cont.) • Algorithm: • Non-Dominated Sorting Genetic Algorithem (NSGA-II). • Tournament selection (starting with size 7 during final 10 generations increases by 20% each generation). • 50 generations. • 500 individuals. • Expression trees are limited to depth of 17 and created with a maximum depth of 8 through Ramped-half-and-half.
Overview • Introduction • Objectives • Multi-objective evolutionary algorithms • Results • Future work
MOEANon-Dominated Sorting Genetic Algorithem • Pareto frontier:
Overview • Introduction • Objectives • Multi-objective evolutionary algorithms • Results • Future work
ResultsAggressiveness– wall collisions avoidance • Fitness = max absolute progress + min wall collisions.
ResultsAggressiveness– max speed & min steering • Fitness = max absolute progress + min wall collisions + min # steering changes.
ResultsAggressiveness– max speed & min steering • Fitness = max absolute progress + min wall collisions + min # steering changes.
ResultsAggressiveness – max speed & min driving • Fitness = max absolute progress + min wall collisions + min # driving changes.
ResultsAggressiveness– max speed & min driving • Fitness = max absolute progress + min wall collisions + min # driving changes.
ResultsAggressiveness– smoothness, avoidance and low speed • Fitness = max absolute progress + min wall collisions + max # driving changes.
ResultsAggressiveness– smoothness, avoidance and low speed • Fitness = max absolute progress + min wall collisions + max # driving changes.
ResultsAggressiveness– max car collisions • Fitness = max absolute progress + max car collisions + min car closeness + min # driving changes.
ResultsAggressiveness– Car collisions • Fitness = max absolute progress + max car collisions + min car closeness + min # steering & driving changes.
ResultsAggressiveness– Opponent weakness Exploitation • Fitness = max absolute progress + max speed + min car closeness + min # steering & driving changes.
Overview • Introduction • Objectives • Multi-objective evolutionary algorithms • Results • Future work
Future work • Prove concept on other game genres.
The End Any Questions ?
The End Thank you ;)