“School of Cognitive”
Back to Papers HomeBack to Papers of School of Cognitive
| Paper IPM / Cognitive / 11400 |
|
||||||||
| Abstract: | |||||||||
|
Concept learning is one of the most important and challenging problems in robotics, artificial intelligence and machine learning. Information is received through robot?s sensors which are practically imposed by many restrictions. This makes it inevitable to consider uncertainty in perception, knowledge representation and decision making. On the other hand, different sources of data might possess different levels of information content for the robot. Therefore, it would be required to fuse information with considering the correct degrees of reliabilities. These reliability rates are generally contextually dependent. In this paper, these understandings are taken into consideration in decision fusion problem in concept learning. Smets? TBM is assumed as the general model of representation and maintaining uncertainty, then according to the complex nature of the obtained constrained optimization problem for learning reliability coefficients of the concept learning agent?s sensors, modeling of the problem is performed for learning by evolutionary algorithms. PBIL is considered here as it outperforms genetic algorithms in many situations. In addition to learning reliabilities using these basic evolutionary methods, an extension of the PBIL in continuous search spaces is proposed then it is shown to be superior to the basic binary PBIL in this problem.
Download TeX format |
|||||||||
| back to top | |||||||||
