Advanced Genetic Programming

1. Advanced Genetic Programming 이상신 SCAI

2. Index • Improving the Speed of GP • Improving Evolvability of Programs • Improving the Power of GP Search

3. Improving the Speed of GP • Run Termination • Parallelization • Parallel Fitness Evaluations • Machine Code Evolution • Stochastic Sampling

4. Cont’d • Run Termination • 실행을 중지하는 여러가지 제한조건 추가 • Intron explosion • Parallelization • Parallel Population • 각 프로세서가 별도의 population 관리 • 인접 프로세서끼리 “boatland”,”migrants”로 통신 • Linear speed-up보다 빠르다.

5. Cont’d • Parallel Fitness Evaluation • 가장 시간이 많이 걸리는 곳은 Evaluation • Evaluate Programs in Parallel • SIMD machine 사용 • Machine Code Evaluation • 프로세서는 기계어를 더 빨리 처리 • 60-200% 속도 향상 • Stochastic Sampling • other fitness selection Technics • Dynamic Subset Selection • Limited Error Fitness

6. 100 different situations Each test takes 500 ms Total 50 programs 각 프로그램이 한번씩만 돌리도록 하자! 한번 진행에 17시간.. Cont’d • Stochastic Sampling

7. Improving the Evolvability of Programs • Modularization • Automatically Defined Functions(ADF) • Encapsulation • Module Acquisition • Adaptive Representation • Loops and Recursion • Memory Manipulation • Strongly Typed Genetic Programming

8. Cont’d • Modularization • Module : lexically contiguous sequence of program statement. • divide and conquor strategy • encapsulation • Crossover building block • Scaling과 inefficiency 문제 해결 • Example • ADF, ADM • Module Acquisition

9. ADF program • Automatically Defined Functions • 구성요소 • result producing set • function set • other ADF • Argument List • Function Definition • Initialize factor • number of function-defining branch • ADF argument # • reference between ADF • function set, terminal set, fitness • ADM defun defun Values ADF0 (ARG0) Values ADF0 * ADF0 (ARG0) * X (ARG0) (ARG0)

10. Cont’d • Module Acquisition • Encapsulation Module Library Argument of module

11. Cont’d • Adaptive Representation • based on heuristic information • make small subtree as candidate of module • these subtree is extracted from population • serve as New functions, subroutines • Epoch : 새로운 candidates for module 집합이 생성될 때까지의 시간 • Subtree의 추출 • 부모와 비교해 가장 높은 Fitness값을 가진 child 프로그램에서 가장 활동적인 부분 • etc • Loop and Recursion • limit number of iteration • distribute execution time according to iteration type • 처음부터 무한루프가 발생하지 않도록 구성

12. Cont’d • Memory Manipulation • assignment of Variables • Complex data structure and Abstract Data Type • GP can handle any data type • evolve its own data type • Cultural GP • Advance by non-genetic information between individuals • Issue • the dissemination speed of genetic information is slow compared to that of cultural information • hence, gene pool is more stable compared to the meme pool • Use common Indexed memory

13. Specify type!! Cont’d • Strongly Typed Genetic Programming N IF S N N AND + S S N N 0 = OR Y 1 N S N S X 1 Z Q

14. Improving the Power of GP Search • Ontology • Adaptive Parsimony Pressure • Chunking • Co-Evolution • Hybrid Approaches

15. Cont’d • Ontology • individual의 lifetime중에 individual의 구성변화 • Structure • carrier of function. Determine function • GP의 문제해결 • How can the search be left unconstrained, whereas the solutions need to be highly constrained through syntax? • How can one use a compact genetic representation that unfolds into a complex individual program? • Is there a way to include the “environment” into process of generating complexity? • Ontology Programming • add self-modification operator

16. Cont’d • Adaptive Parsimony Pressure • Similar to human’s problem solving • first find a solution • next simply solution and make it more elegant Error produced by individual i Complexity of individual i generation

17. Cont’d • Chunking • divide big fitness case into small chunks • ex) image process • Co-Evolution • 부모와 자식 세대가 서로 fitness를 평가.

18. Cont’d • Hybrid Approach • GP + other search algorithm • general search algorithm • special purpose search algorithm