Page 256 - Mechatronics for Safety, Security and Dependability in a New Era

Ch48-I044963.fm Page 240 Tuesday, August 1, 2006 4:04 PM
Ch48-I044963.fm
240
240 Page 240 Tuesday, August 1, 2006 4:04 PM

Fs
" Fc
Fm
1st
2nd
3rd
(a) Navigation task 50 100 150 200 250
step
(c) Selected feature extractors
ftepO

(b) Resultant behavior (d) Selected subsets of instance
Figure 3: Task and experimental results

effective feature extractors.

To verify the subset of instances, we performed the same experiment except for the procedure to select
the subsets. In this experiment, the robot always selected all instances. In the result, the average
number of selected feature extractors per step is 1.97, which is larger than the result of Figure 3. This
means that the robot spent much more time for action decision at each step. Hence, the robot
effectively decides the action using a portion of the instances.

CONCLUSION
This paper has proposed a method in which a robot learns to select image feature extractors generated
by itself according to a task-relevant criterion. A portion of supervised data which gives the local
information of the task makes the selection of feature extractors more effective. In the proposed
method, a robot can accomplish more complicated tasks using multiple feature extractors. Our future
work is to verify the extent of effectiveness of the proposed method.

References

McCallum, A. (1996). Learning to Use Selective Attention and Short-Term Memory in Sequential
Tasks. Proceedings of International Conference on Simulation of Adaptive Behavior, 315-324.

Minato, T. and Asada, M. (2003). Towards Selective Attention: Generating Image Features by
Learning a Visuo-Motor Map. Robotics and Autonomous Systems 45, 211-221.

Mitsunaga, N. and Asada, M. (2000). Observation Strategy for Decision Making based on
Information Criterion. Proceedings of International Conference on Intelligent Robots and Systems,
1038-1043.

Vlassis, N., Bunschoten, R., and Krose, B. (2001). Learning Task-Relevant Features from Robot
Data. Proceedings of International Conference on Robotics and Automation, 499-504.