DocumentCode :
2444642
Title :
A new methodology for reducing brittleness in genetic programming
Author :
Moore, Frank W. ; Garcia, Oscar N.
Author_Institution :
Wright State Univ., Dayton, OH, USA
Volume :
2
fYear :
1997
fDate :
14-18 Jul 1997
Firstpage :
757
Abstract :
Genetic programming systems typically use a fixed training population to optimize programs according to problem-specific fitness criteria. The best-of-run programs evolved by these systems frequently exhibit optimal (or near-optimal) performance in competitive survival environments explicitly represented by the training population. Unfortunately, subsequent performance of these programs is often less than optimal when situations arise that were not explicitly anticipated during program evolution. This paper describes a new methodology which promises to reduce the brittleness of best-of-run programs evolved by genetic programming systems. Instead of using a fixed set of fitness cases, the new methodology creates a new set of randomly-generated fitness cases prior to the evaluation of each generation of the evolutionary process. A genetic programming system that evolves optimized maneuvers for an extended 2D pursuer/evader problem was modified for this study. The extended 2D pursuer/evader problem is a competitive zero-sum game in which an evader attempts to escape a faster, more agile pursuer by performing specific combinations of thrusting and turning maneuvers. The pursuer uses the highly effective proportional navigation algorithm to control its trajectory towards the evader. The original genetic programming system used a fixed training set of pursuers. Each of these pursuers was uniquely identified by two parameters: the initial distance from pursuer to evader, and the angle that the velocity vector of the evader makes relative to the pursuer/evader line-of-sight at the time the pursuer is launched. The modified system implemented for this project was identical to the original system, except that it used random distances and angles to create a new set of fitness cases prior to each generation of the genetic programming run. Best-of-run programs were independently evolved using fixed and randomly-generated fitness cases. These programs were subsequently tested against a large, representative fixed population of pursuers to determine their relative effectiveness. This paper describes the implementation of both the original and modified systems, and summarizes the results of these tests
Keywords :
algorithm theory; competitive algorithms; differential games; genetic algorithms; learning (artificial intelligence); missile guidance; best-of-run programs; brittleness reduction methodology; competitive survival environments; competitive zero-sum game; differential games; evolutionary programming; extended 2D pursuer/evader problem; fixed training population; genetic programming; missile guidance; optimal strategy; optimized maneuvers; parse trees; problem-specific fitness criteria; proportional navigation algorithm; pursuer-evader initial distance; pursuer/evader line-of-sight; randomly-generated fitness cases; velocity vector angle; Aggregates; Genetic mutations; Genetic programming; Knowledge acquisition; Navigation; Optimization methods; Proportional control; System testing; Turning;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Aerospace and Electronics Conference, 1997. NAECON 1997., Proceedings of the IEEE 1997 National
Conference_Location :
Dayton, OH
Print_ISBN :
0-7803-3725-5
Type :
conf
DOI :
10.1109/NAECON.1997.622725
Filename :
622725
Link To Document :
بازگشت