In recent years, the multitask optimization problem has attracted attention as a new problem class to which evolutionary computation is applied. The multitask optimization problem is an optimization problem that aims to solve multiple tasks in parallel. In conventional optimization problems, a solver is prepared for each task and tuning is required. However, in multitask optimization problems, the tuning of one solver is enough, so the burden on the designer can be reduced. In addition, by solving multiple optimization problems in parallel, it becomes possible to share knowledge among tasks, so that efficient search can be expected. In a single-objective multitask optimization problem, inter task crossover has a positive effect between similar tasks and a negative effect between dissimilar tasks. The conventional multitask optimization methods show good performance in similar multiple task optimization. However, when tasks are not similar, these conventional methods decrease the interaction such as inter task crossover. Therefore, the search performances are not different from that of the single-objective optimization methods independently applied to each task. In this paper, we attempt to improve the similarity between tasks by rearranging the notation of design variables for tasks that are not similar. This paper shows that the search performance can be improved in the widely used two-task benchmark multitask optimization problems.
In the present work, we propose some modifications on multiple constraint ranking (MCR) to improve its performance in handling constraints in engineering design optimization with evolutionary algorithm. Considering that the search from both feasible and infeasible regions is more efficient, the modifications are proposed so that MCR can adaptively conduct more exploration in infeasible region. Based on investigation in a car structure design optimization problem, some of the proposed modifications have proven to be significantly effective in enhancing the convergence performance towards constrained optimum in this particular problem compared with the original MCR. We discover that the modifications which significantly improve the convergence performance produce more variety of infeasible individuals compared with MCR. Also, many of infeasible individuals produced by those modifications have better objective values than the feasible ones while we hardly observe similar occurrence in MCR. These two factors might make those modifications have better interaction between feasible and infeasible regions, thus produce convergence improvement on MCR.