Creating a Multi-Purpose First Person Shooter Bot with Reinforcement Learning
12
2009/10/08 1 Creating a Multi-Purpose First Person Shooter Bot with Reinforcement Learning IEEE Symposium On Computational Intelligence and Games pp. 143-150 Dec. 2008 Authors: M. McPartland and M. Gallagher Presented by Chien-Erh Huang
description
Creating a Multi-Purpose First Person Shooter Bot with Reinforcement Learning. IEEE Symposium On Computational Intelligence and Games pp. 143-150 Dec. 2008 Authors: M. McPartland and M. Gallagher Presented by Chien-Erh Huang. 1. 2. 3. 4. 5. Introduction. Background. Method. Results. - PowerPoint PPT Presentation
Transcript of Creating a Multi-Purpose First Person Shooter Bot with Reinforcement Learning
2009/10/08 1
Creating a Multi-Purpose First Person Shooter Bot with Reinforcement Learning
IEEE Symposium On Computational Intelligence and Gamespp. 143-150 Dec. 2008Authors: M. McPartland and M. GallagherPresented by Chien-Erh Huang
2
Outline
Introduction1
Background2
Method3
Results4
CONCLUSIONSConclusions5
3
Introduction(1/2)
Bot artificial intelligence (AI) in first person shooter (FPS) games generally comprise of path planning, picking up items, and combat.
Traditionally, hard-coded methods such as state machines, rule based systems, and scripting are used for bot AI in commercial games. The problem with these methods is that they are static, can be hard to expand, and time is needed to hand tune parameters.
This paper investigates several methods based on reinforcement learning (RL) to create a multi-purpose bot AI with the behaviors of navigation, item collection and combat.
4
Introduction(2/2)
This paper expands on previous work where low-level FPS bot controllers were trained using RL. The aim of this paper is to use the previously trained controllers and combine them to produce bots with a multi-purpose behavior set.
Three different types of RL will be compared to investigate the differences in statistics and behaviors.
● Hierarchical RL
● Rule based RL
● RL
5
BACKGROUND(1/2)
hierarchical
6
BACKGROUND(2/2)
Sarsa
Q=policye=eachs=statea=actionδ=variabler=reward =decay parameter =learning rate
7
METHOD(1/2)
Yellow represents the Bots spawn positions, red and green represent item points.
arena map maze map
8
METHOD(2/2)
All three RL bots were trained for 5000 iterations in both the arena and maze environments.
9
RESULTS(1/2)
10
RESULTS(2/2)
11
CONCLUSIONS
Both the arena and maze maps showed similar trends in combat and navigation statistics indicating the robustness of the RL controllers.
The hierarchical RL and rule based RL controllers performed significantly better than the flat RL controller in the combat and navigation skills.
The hierarchical RL bot performed best in the shooting accuracy objective, outperforming all other bots in the experiment.
The rule based RL bot performed slightly better in the other objectives than the hierarchical RL bot.
2009/10/08 12
