| 研究課題/領域番号 |
25330253
|
|---|
| 研究種目 |
基盤研究(C)
|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
福永 ALEX 東京大学, 総合文化研究科, 准教授 (90452002)
|
|---|
| 研究期間 (年度) |
2013-04-01 – 2017-03-31
|
| キーワード | 進化計算 / 探索アルゴリズム |
|---|
| 研究概要 |
Differential Evolution is a simple, but effective approach for numerical optimization. Since the search efficiency of DE depends significantly on its control parameter settings, there has been much recent work on developing self-adaptive mechanisms for DE. We propose a new, parameter adaptation technique for DE which uses a historical memory of successful control parameter settings to guide the selection of future control parameter values. We showed that the experimental results show that a DE using our success-history based parameter adaptation method is competitive with the state-of-the-art DE algorithms.
We investigated a large-scale, empirical evaluation of a Random, Heterogeneous Island-Model (RHIM) for evolutionary algorithms (EAs), where the control parameter values are independently, randomly assigned for each island that has recently been proposed by Gong and Fukunaga as a method for configuring island-model evolutionary algorithms in situations where it is not possible to expend the resources to carefully tune control parameters for a particular application. We applid RHIM to standard DE, JADE (an adaptive DE), and real-coded genetic algorithms, and showed that the search efficiency of RHIM is compared to manual tuning of parameter settings for each benchmark problem.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
One of the goals of this Kakenhi project is the development of large-scale heterogeneous algorithm portfolio construction techniques.
This year, we investigated a simple, baseline method for constructing portfolios where a portfolio consists of different instances of a parameterized search algorithm assigned to each processor in a distributed system (where the parameter values are randomly, uniformly selected from a set of reasonable ranges).
We compared this simple method to several regimes of hand-tuning of parameters for a wide range of evolutionary algorithms, on a broad range of benchmark problems. Our results indicate that the simple, island-based randomized parameter setting is comparable to the kinds of "shallow" attempts at hand tuning that are feasible in practice. Furthermore, we showed that as the number of processors increased, the method is performs increasingly well.
Thus, we have established a new simple, baseline for heterogeneous portfolios that should serve as the basis for future work in this area.
|
| 今後の研究の推進方策 |
In 2014, we will continue our investigation of the simple, randomized portfolio method we investigated this year (Randomized Heterogeneous Island Model). In addition, after initial consideration of portfolio construction for several classes of problems, including real-valued function optimization and domain-independent planning, we believe that in order to construct portfolios that achieve state-of-the-art performance in these problems, we needed to develop new algorithms for these problems that will be included in the large-scale algorithm portfolios that we seek to develop. Therefore, in 2014, we will develop base-level search algorithms for real-valued function optimization, as well as algorithms for domain-independent planning.
Furthermore, we are evaluating the characteristics of problem instances included in common benchmark problem sets for real-valued function optimization and domain-independent planning. This is expected to yield insights into the types of algorithm combinations that will result in successful portfolios.
In 2015-7, we will continue these lines of research by seeking to integrate the new base level algorithms developed in 2014 into high-performance portfolios and evaluating them on standard benchmark problem sets.
|
| 次年度の研究費の使用計画 |
In 2013, we purchased a server/workstation for computational experiments, which cost 751,065yen. The remaining amount (under 50,000 yen) is therefore being carried over for the next fiscal year.
In 2014, we may purchase another server for computational experiments, depending on the condition of our existing servers. The amount carried over from 2013 will be used to fund part of this purchase, or for travel costs or office equipment/supplied.
|
