In this thesis we consider two improve versions of the differential evolution. First, the differential evolution with growing structure is considered. If a particle exploring the optimum fall into a local optimal solution then a new particle is born and particles share the position imformation. And particles can help escape from the trap of the local optimal solution. Performing numerical experiments for basic benchmarks, the algorithm efficiency is evaluated in several key indicators such as typical search example, success rate, the number of interation and number of particle. Next, insensitive differential evolution for multi-solution problems is considered. The Algorithm consists of global and local searches. In the global search, the algorithm tries to construct local sub-regions (LSRs) each of which includes either solution. In the local search, the algorithm operates on all the LSRs in parallel and tries to find all the solutions. The algorithm has a key parameter that controls the algorithm insensitivity. If the insensitivity is suitable, the algorithm can construct all the LSRs before trapping into either solution and can find all the solutions. Performing basic numerical experiments where parameters are adjusted by trial-and-errors, basic performance of the algorithm is investigated.

114 工学部 (旧) ・工学研究科 ＞ 学位論文 ＞ 修士論文