2003 Fiscal Year Final Research Report Summary
Fundamental Research on Advanced Evolutionary and Adaptive Systems
| Project/Area Number |
13480089
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Tokyo Institute of Technology |
|---|
Principal Investigator |
KOBAYASHI Shigenobu Tokyo Institute of Technology, Interdisciplinary School of Science and Technology, Professor, 大学院・総合理工学研究科, 教授 (40016697)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
KIMURA Hajime Tokyo Institute of Technology, Interdisciplinary School of Science and Technology, Research Associate, 大学院・総合理工学研究科, 助手 (40302963)
|
|---|
| Project Period (FY) |
2001 – 2003
|
| Keywords | Evolutionary Computation / Real-coded Genetic Algorithms / UV Structure / k-tablet Structure / Reinforcement Learning / Actor-Critic / four-legged Walking Robot / Distributed Reinforcement Learning |
|---|
| Research Abstract |
A.Research Results on Evolutionary Computation
・We proposed a robust real-coded GA using the combination of two crossovers, UNDX-m and EDX. It can deal both ridge-structure function whose dimension reaches more than hundreds and multi-peak functions.
・We proposed a new evolutionary algorithm called ANS (Adaptive neighboring Search) with a crossover-like mutation to optimize high dimensional deceptive multimodal functions.
・We found a hypothesis call "UV-phenomenon" which explains failures of global search by GA. It suggests UV-structures as hard landscape structures that will cause the UV-phenomenon.
B.Research Results on Reinforcement Learning
・We proposed a new crossover LUNDX-m which uses only in-dimensional latent variables. LUNDX-m can treat with high-dimensional ill-scaled structures called k -tablet structure.
・In multi-agent reinforcement learning systems, it is important how to share a reward among all agents. We derived the necessary and sufficient condition to preserve the rationality to realize cooperative behaviors.
・We proposed a policy function representation that consists of a stochastic binary decision tree. We applied it to an actor-critic algorithm for the problems that have enormous similar actions.
・We investigated a reinforcement learning of walking behavior for a four-legged robot. We presented. a new actor-critic algorithm, in which the actor selects a continuous action from its bounded action space by using the normal distribution.
・We proposed a new simulation-based distributed reinforcement learning approach that solves large planning problems under uncertain environment. We applied it to real sewerage control systems.
|
