@phdthesis{oai:kyutech.repo.nii.ac.jp:00000793,
 author = {Muslim, Muhammad Aziz},
 month = {2008-09-25},
 note = {1 Introduction||2 Overview of Related Studies||3 Task Segmentation using mnSOM and Clustering||4 Formation of Graph-based Maps||5 Experimental Results on Task Segmentation and Clustering||6 Experimental Results on the Formation of Graph basedMaps||7 Conclusions and Discussions||Bibliography, A new approach in Artificial Intelligence (AI), which focuses on agent’s interaction with the world, is expected to solve difficulties in the classical AI. The interaction leads an agent to exhibit emergent behaviors, which are not preprogrammed by a designer. This is what biological agents (i.e. animals and humans) do in their daily life. This dissertation aims at finding mechanisms necessary for this. A mobile robot is used as a test bed for this purpose. The real world is completely different from a virtual world frequently used in the classical AI. The real world is always subject to complexity, noise, and nonlinearity. Information on the real world is often spatio-temporal in nature, and is hard to do information processing in real time. Solving real world problems often leads to unsatisfactory results due to its inherent difficulties. One promising approach to this is to segment the spatio-temporal information into meaningful elements. The purpose of the present thesis is to segment the world and to form a graph-based map for efficient processing. Task segmentation in navigation of a mobile robot based on sensory signals is important for realizing efficient navigation, hence attracted wide attention. In this research, a new approach to segmentation in a mobile robot by a modular network SOM (mnSOM) is proposed. In a mobile robot, the standard mnSOM is not applicable as it is, because it is based on an assumption that class labels are known a priori. In a mobile robot, however, only a sequence of data without segmentation is available. Hence, we propose to decompose it into many subsequences, supposing that a class label does not change within a subsequence. Accordingly, training of mnSOM is done for each subsequence in contrast to that for each class in the standard mnSOM. The resulting mnSOM demonstrates segmentation performance of 94.05% for a novel dataset based on an unrealistic assumption that winner modules corresponding to subsequences in the same class share the same label. Since this is not at all practical, the current study proposes segmentation without this unrealistic assumption. Firstly, the conventional hierarchical clustering is applied to the resulting mnSOM. Without the above unrealistic assumption, its segmentation performance deteriorates by only 1.2%. Hierarchical clustering assumes that the distances between any pair of modules are provided with precision, but this is not the case in mnSOM. Accordingly, this is followed by a clustering based on only the distance between spatially adjacent modules with modification by their temporal contiguity. This clustering with spatio-temporal contiguity provides superior performance to the conventional hierarchical clustering. Based on the resulting mnSOM, a graph-based map is formed. Due to stochastic character of sensory-motor information, I propose to use Hidden Markov models (HMMs) instead of a deterministic method. Given a sequence of data, mnSOM produces sequence of labels, which may includes erroneous ones due to noise. HMMs are employed for better estimates of labels. Finally, from the resulting sequence of labels, L-junctions and T-junctions are located, and are used as nodes for constructing a graph-based map. For comparative study, vector quantization of sensory-motor signals is also tried. The resulting HMMs based on the quantized data also generate a graph-based map. The resulting graph-based map also contributes to goal seeking. Simulation result shows that the resulting graph-based map is efficient for goal seeking, since it is not necessary to construct a new map every time the environment changes., 九州工業大学博士学位論文 学位記番号:生工博甲第86号 学位授与年月日:平成20年3月25日, 平成19年度},
 school = {九州工業大学},
 title = {Segmentation of Environment and Formation of Graph-based Maps in Mobile Robots},
 year = {}
}