CC BY
35
Известия ТРТУ
Тематический выпуск
ТЕЗИСЫ ДОКЛАДОВ НАУЧНО-ТЕХНИЧЕСКОЙ КОНФЕРЕНЦИИ «ИНТЕЛЛЕКТУАЛЬНЫЕ САПР-97»
yflK 658.512
S.Dergatchev, V.Kureichik Allocation problem using genetic algorithm
Abstract: This paper describes a novel approach to allocation problem in high-level synthesis using genetic algorithms(GA).This approach is different from a previous attempt using GA[1] in many respects. Our contributions include: a new chromosomal representation for two problems of allocation; and two novel crossover operators to generate legal allocations. Operation hardware allocation is the important phase in the synthesis of circuits from behavioral descriptions. Several optimization techniques can be used for this purpose such as simulation annealing[2] and integer programming^]. Genetic algorithms is another promising global optimization technique[4]. It works by emulating the natural process of evolution as a means of progressing toward the optimum. The algorithm starts with a population which consists of several solutions to the optimization problem. A member of population is called an individual. A fitness value is associated with each individual. Each solution in the population or an individual is encoded as a string of symbols. These symbols are known as genes and the solution string is called a chromosome. The values taken by genes arc called alleles. Several pair of individuals(parents) in the population mate to produce offsprings by applying the genetic operator crossover. Selection of parents is done by repeated use of choice tunction. A number of individuals and offsprings are passed to a new generation such that the number of individuals in the new population is the same as old population. A selection function determines which strings form the population in the next generation. Each surviving string undergoes mutation and inversion with a specified probability. Fitness scaling is used to avoid premature convergence. One method is linear scaling[4]. Linear scaling runs into problems in later runs of the genetic algorithm when most of the fitness values are close to each other and some lethal members have very low fitness values. Since we want to combine scheduling and allocation into one optimization problem, the coding has to reflect this. This can be done only to certain extent as finding an encoding for all the parameters is nearly impossible as there are too many constraints. Each gene has three values control-step number, functional unit number, and the number of the functional unit input to the left variable of the operation is assigned. Genetic allocation is tested on various benchmarks. The results are compared with allocation using tabu search and simulated evolution.
References
1. N. Wehn and other, A novel allocation approach for datapath synthesis based on genetic paradigm, In IFIP Working Conference on Logic and Architecture Synthesis, Paris, pages 47-56, 1990.
2. S. Devadas and A. Newton., Algorithms for hardware allocation in data path synthesis, IEEE Transactions on Computer-Aided Design, 8(7):768-781, 1989
3. M. Balakrishnan and P. Marwedel, Integrated allocation and binding: A synthesis approach for design space exploration, in Proceedings of the 26lh Design Automation Conference, pp. 68-74, 1989.
4. David E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Publishing Company, Inc., 1989.
