2003 Fiscal Year Final Research Report Summary
Learning, Understanding, Analysis and Re-use Of Robot's Moving Strategies
| Project/Area Number |
14580426
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | THE UNIVERSITY OF AIZU |
|---|
Principal Investigator |
ZHAO Qiangfu THE UNIVERSITY OF AIZU, School of Computer Science and Engineering, Professor, コンピュータ理工学部, 教授 (90260421)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
LIU Yong THE UNIVERSITY OF AIZU, School of Computer Science and Engineering, Associate Professor, コンピュータ理工学部, 助教授 (60325967)
|
|---|
| Project Period (FY) |
2002 – 2003
|
| Keywords | Mobile robots / Neural networks / Evolutionary learning / Decision Trees / Neural network trees / Rule extraction / Rule understanding / Rule re-use |
|---|
| Research Abstract |
In these two years, we mainly studied the following two topics :
(1)Learning and understanding of neural network robot controllers : In designing an autonomous robot, it is not practical to tell the robot the correct answers for all possible situations. In such cases, evolutionary learning or reinforcement learning is considered more efficient. In this research we adopted evolutionary learning. The robot used here is the well-known mini-robot Khepera. The purpose is to (a) Evolve a neural network that can control the robot go approach to a goal (a light source) or go around freely while avoiding obstacles ; (b) Extract rules from the neural network controller ; and (c) Analyze the rules, and re-use the useful rules to accelerate the evolutionary learning of other robot controllers. So far we have obtained some good results for (a) and (b), and we have published papers in international conferences. For (c), however, there are still some open problems. The main obstacle in analyzing the r
… More
ules is that the rules extracted from the neural network controllers are often too complex to understand. Currently, we proposed some methods to solve this problem, and we are doing further research to confirm these methods.
(2)Learning of interpretable and comprehensible neural network controllers : To extract rules from a trained neural network is in general a NP-complete problem. To solve this problem more efficiently, we proposed the neural network tress (NNTrees). An NNTree is a decision tree with each non-terminal node containing an expert neural network (ENN). If we limit the number of features used in each ENN, we based genetic algorithm for designing NNTrees that are both interpretable and comprehensible. From the experiments we have the following conclusions : (a) The NNTrees obtained are very easy to interpret, and (b) The performance of the trees is not decreased. However, the time complexity for design itself is increased. Therefore, it is still not practical to use NNTrees for robot control. Less
|
Research Products
(12 results)
