Advanced Genetic Programming SCAI Index t Improving the

Advanced Genetic Programming 이상신 SCAI

Index t Improving the Speed of GP t Improving Evolvability of Programs t Improving the Power of GP Search

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

Cont’d t Run Termination i실행을 중지하는 여러가지 제한조건 추가 i. Intron explosion t Parallelization i. Parallel Population 프로세서가 별도의 population 관리 u 인접 프로세서끼리 “boatland”, ”migrants”로 통신 u Linear speed-up보다 빠르다. u각

Cont’d t Parallel Fitness Evaluation i 가장 시간이 많이 걸리는 곳은 Evaluation i Evaluate Programs in Parallel i SIMD machine 사용 t Machine Code Evaluation i 프로세서는 기계어를 더 빨리 처리 i 60 -200% 속도 향상 t t Stochastic Sampling other fitness selection Technics i Dynamic Subset Selection i Limited Error Fitness

Cont’d t Stochastic Sampling 100 different situations 각 프로그램이 한번씩만 돌리도록 하자! Each test takes 500 ms Total 50 programs 한번 진행에 17시간. .

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

Cont’d t Modularization i. Module : lexically contiguous sequence of program statement. idivide and conquor strategy iencapsulation u Crossover building block i. Scaling과 inefficiency 문제 해결 i. Example u ADF, ADM u Module Acquisition

ADF t Automatically Defined Functions t 구성요소 i result producing set i function set other ADF u Argument List u Function Definition program defun Values u t ADF 0 (ARG 0) ADF 0 Initialize factor i number of function-defining branch i ADF argument # i reference between ADF i function set, terminal set, fitness t Values ADM * (ARG 0) ADF 0 * X (ARG 0)

Cont’d t Encapsulation t Module Acquisition Module Library Argument of module

Cont’d t Adaptive Representation i based on heuristic information make small subtree as candidate of module u these subtree is extracted from population u s serve as New functions, subroutines i Epoch : 새로운 candidates for module 집합이 생성될 때까 지의 시간 i Subtree의 추출 u t 부모와 비교해 가장 높은 Fitness값을 가진 child 프로그램에 서 가장 활동적인 부분 etc i Loop and Recursion u limit number of iteration u distribute execution time according to iteration type u 처음부터 무한루프가 발생하지 않도록 구성

Cont’d t Memory Manipulation iassignment of Variables i. Complex data structure and Abstract Data Type u GP can handle any data type u evolve its own data type i. Cultural GP u Advance by non-genetic information between individuals u Issue s s u Use 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 common Indexed memory

Cont’d t Strongly Typed Genetic Programming N IF S N AND S X N + S = N Specify type!! N Y OR N S 1 Z S Q N 1 0

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

Cont’d t Ontology iindividual의 lifetime중에 individual의 구성변화 i. Structure u carrier of function. Determine function i. GP의 문제해결 u How can the search be left unconstrained, whereas the solutions need to be highly constrained through syntax? u How can one use a compact genetic representation that unfolds into a complex individual program? u Is there a way to include the “environment” into process of generating complexity? t Ontology Programming iadd self-modification operator

Cont’d t Adaptive Parsimony Pressure Error produced by individual i Complexity of individual i generation i. Similar to human’s problem solving u first find a solution u next simply solution and make it more elegant

Cont’d t Chunking i divide big fitness case into small chunks i ex) image process t Co-Evolution i 부모와 자식 세대가 서로 fitness를 평가.

Cont’d t Hybrid Approach i. GP + other search algorithm u general search algorithm u special purpose search algorithm