Source-linked AI summary
Niching Agents in The Core
Gary B. Parker, Jim O'Connor, John Asaro
TL;DR
This study asks whether The Core’s agents can be niched to subsets of the Xpilot environment rather than the whole map. It restricts respawns to each agent’s starting quadrant and reports success for niched agents, especially those evolving in less obstacle-dense quadrants.
Problem
The study investigates whether The Core can evolve agents specialized to subsets of an environment rather than the entire map.
Source: §1.3 · p. 4Method
The Core co-evolves Xpilot controllers through local interactions, with agents restricted to respawn within their initial quadrant.
Source: §1.3 · p. 4Results
Agents from less obstacle-dense quadrants outperformed visiting agents in their native environments and dominated opponents in foreign quadrants.
Source: §3 · p. 8Takeaways & Limitations
The authors report that The Core can niche agents to subsets of a larger environment and that its controllers can exhibit emergent obstacle detection while focusing on combat.
Source: §4 · pp. 8, 10
Abstract
from arXiv · showhide
The Core is a unique competitive co-evolution algorithm that allows agents to evolve autonomous control without utilizing a traditional fitness function. The agents evolve via local interactions through tournament selection, crossover, and mutation, producing offspring by evolving better controllers. Previous works have shown The Core's ability to evolve agents capable of combat and navigation in the Xpilot video game. This research expands upon that premise by niching agents to specific subsets of the original environment The Core was tested in. Our results demonstrate the niched agents capacity for success over agents niched to the entire system and agents niched to different sub-environments.
1 Introduction
Evolutionary algorithms have been used to develop adaptive game agents, and Xpilot-AI offers a dynamic multiplayer testbed. This study uses The Core with quadrant-restricted respawns to investigate specialized behavior in distinct environments.
1.1 Evolutionary Algorithms in Games
- Evolutionary algorithms have been applied to complex, dynamic games, including Xpilot-AI, which supports many agents and sparse rewards.Prior work also describes agents competing against human and computational opponents, including NERO’s use of rtNEAT to adapt behavior.Source: §1.1 · p. 1
1.2 Environment
- Xpilot is a multiplayer, two-dimensional space-combat game where players pilot ships, and this study limits agent controls to turning, thrusting, and shooting.The map used here has four quadrants with distinct obstacle sizes and densities.Source: §1.2 · pp. 2–3
1.3 The Core
- The Core co-evolves autonomous Xpilot controllers by propagating beneficial behavioral traits from successful agents to nearby individuals.The algorithm is intended to support navigation and combat in complex terrain.Source: §1.3 · p. 4
- Unlike the earlier setup that allowed respawning anywhere on the map, this study restricts each agent’s respawn to its initial quadrant.The earlier generalized implementation is attributed to.Source: §1.3 · p. 4
- Restricting agents to local terrain is called niching and is intended to foster behaviors adapted to each quadrant’s conditions.The paper links quadrant-specific terrain to specialized navigation, combat, and obstacle-avoidance strategies.Source: §1.3 · p. 4
2 Methodology
The methodology combines a multi-loop cyclic genetic algorithm with The Core’s competitive reproduction process. It divides the population among map quadrants so agents evolve in locally distinct terrain.
2.1 Multi-Loop Cyclic Genetic Algorithm
- Multi-loop cyclic genetic algorithms use multiple loops for varying state spaces, switching between loops when specified conditionals are met.The method builds on cyclic genetic algorithms originally developed for repetitive hexapod gaits and later extended for more complex state spaces.Source: §2.1 · p. 4
2.2 The Core Algorithm
- Each controller uses 12 loops of eight 10-bit genes; condition-triggered jumps select loops for specific scenarios, while action genes encode agent behavior.Loops are grouped into special, speed-, direction-, and distance-based categories.Source: §2.2 · p. 5
- When an agent destroys another with a projectile, the defeated agent respawns after 32 frames with a chromosome crossed over from both agents.The Core uses this competitive interaction instead of a fitness function.Source: §2.2 · p. 5
- Agents first spawn as children and can spread their genes only after reaching three seconds of age.The paper calls this threshold the agent’s “age of adolescence.”Source: §2.2 · p. 5
2.3 Niching
- The population of 120 is split into four groups of 30, each restricted to one quadrant with distinct obstacle patterns.The paper expects less obstacle-dense areas to favor combat strategies, while denser areas may require more survival and wall-avoidance tactics.Source: §2.3 · p. 7
3 Results
The evaluation compared quadrant-native agents with agents trained across the whole environment and tested whether native agents could compete after transfer to adjacent quadrants. Agents from lower-obstacle-density quadrants outperformed visiting agents at home and dominated opponents in foreign quadrants.
3 Results
- Sixteen quadrant-native agents were matched against 16 globally niched agents, with each matchup repeated 10 times for 12 hours.Agents were evolved for 72 hours; friendly fire was disabled and teammates were excluded as targets.Source: §3 · p. 7
- The transfer test paired 16 agents native to each quadrant against 16 agents from the preceding quadrant.The pairings were Q1–Q4, Q2–Q1, Q3–Q2, and Q4–Q3; Figure 4 presents the test layout.Source: §3 · p. 8, Figure 4 · p. 8
- Agents from lower-obstacle-density quadrants 1 and 3 outperformed visitors at home and dominated opponents after transfer to foreign quadrants.The authors suggest adaptation to less obstacle-dense environments may let these agents prioritize combat over obstacle avoidance.Source: §3 · p. 8
4 Conclusions
The paper reports that The Core can evolve Xpilot controllers for navigation and combat, and extends this approach to agents niched to subsets of the environment. It also describes flexible decision-making, parallel evolution, and possible future tests in other environments.
4 Conclusions
- The work demonstrates a solution for niching agents to subsets of a larger environment, building on The Core's Xpilot navigation and combat results.The paper describes the combined task as challenging and says it can be effectively solved with a multi-loop CGA and The Core.Source: §4 · p. 8
- The Core's dynamic chromosomes allow combat-focused agents to exhibit emergent obstacle detection by turning toward the closest enemy ship.The authors contrast this flexibility with expert-system agents that might always select obstacle detection under specified thresholds.Source: §4 · p. 10
- The Core evolves over 100 agents in parallel and niches large populations in relatively short amounts of time, whereas larger populations traditionally increase input space and evolution time.The authors present this as a possible advantage for similarly challenging game environments.Source: §4 · p. 10
- Future work could vary maps, agent counts, play rules, and game counts, or apply The Core to simulated racing and more complex games.These are proposed extensions, not evaluated results in this work.Source: §4 · p. 10