Evolutionary Improvement of Programs

TY - JOUR

T1 - Evolutionary Improvement of Programs

AU - White, David Robert

AU - Arcuri, Andrea

AU - Clark, John A.

PY - 2011/8

Y1 - 2011/8

N2 - Most applications of genetic programming (GP) involve the creation of an entirely new function, program or expression to solve a specific problem. In this paper, we propose a new approach that applies GP to improve existing software by optimizing its non-functional properties such as execution time, memory usage, or power consumption. In general, satisfying non-functional requirements is a difficult task and often achieved in part by optimizing compilers. However, modern compilers are in general not always able to produce semantically equivalent alternatives that optimize non-functional properties, even if such alternatives are known to exist: this is usually due to the limited local nature of such optimizations. In this paper, we discuss how best to combine and extend the existing evolutionary methods of GP, multiobjective optimization, and coevolution in order to improve existing software. Given as input the implementation of a function, we attempt to evolve a semantically equivalent version, in this case optimized to reduce execution time subject to a given probability distribution of inputs. We demonstrate that our framework is able to produce non-obvious optimizations that compilers are not yet able to generate on eight example functions. We employ a coevolved population of test cases to encourage the preservation of the function's semantics. We exploit the original program both through seeding of the population in order to focus the search, and as an oracle for testing purposes. As well as discussing the issues that arise when attempting to improve software, we employ rigorous experimental method to provide interesting and practical insights to suggest how to address these issues.

AB - Most applications of genetic programming (GP) involve the creation of an entirely new function, program or expression to solve a specific problem. In this paper, we propose a new approach that applies GP to improve existing software by optimizing its non-functional properties such as execution time, memory usage, or power consumption. In general, satisfying non-functional requirements is a difficult task and often achieved in part by optimizing compilers. However, modern compilers are in general not always able to produce semantically equivalent alternatives that optimize non-functional properties, even if such alternatives are known to exist: this is usually due to the limited local nature of such optimizations. In this paper, we discuss how best to combine and extend the existing evolutionary methods of GP, multiobjective optimization, and coevolution in order to improve existing software. Given as input the implementation of a function, we attempt to evolve a semantically equivalent version, in this case optimized to reduce execution time subject to a given probability distribution of inputs. We demonstrate that our framework is able to produce non-obvious optimizations that compilers are not yet able to generate on eight example functions. We employ a coevolved population of test cases to encourage the preservation of the function's semantics. We exploit the original program both through seeding of the population in order to focus the search, and as an oracle for testing purposes. As well as discussing the issues that arise when attempting to improve software, we employ rigorous experimental method to provide interesting and practical insights to suggest how to address these issues.

UR - http://www.scopus.com/inward/record.url?scp=79961024706&partnerID=8YFLogxK

U2 - 10.1109/TEVC.2010.2083669

DO - 10.1109/TEVC.2010.2083669

M3 - Article

SN - 1089-778X

VL - 15

SP - 515

EP - 538

JO - IEEE Transactions on Evolutionary Computation

JF - IEEE Transactions on Evolutionary Computation

IS - 4

M1 - 5688317

ER -