Search

Partial observability, or the inability of an agent to fully observe the state of its environment, exists in many real-world problem domains. However, most cognitive architectures do not have a theoretical foundation that allows for a systematic approach to handling partial observability. To address this issue, in this paper, we propose a novel cognitive architecture based on the theory of partially observable Markov decision processes (POMDPs). We focus on one of the most fundamental cognitive architectures, the blackboard architecture, and re-formulate it using the POMDP framework so that the agent can automatically decide how to use its information processing modules in partially observable environments. We propose a two-layer POMDP model for the blackboard architecture in which one layer models the dynamics of the environment and the other layer models the dynamics of the blackboard and knowledge sources within the agent. We also provide a planning method to find the optimal actions for the two-layer POMDP model based on the partially observable upper confidence bounds applied to trees (PO-UCT) algorithm. We validate the proposed architecture in both numerical and real-robot experiments.

This paper describes design and development of a system that supports continuous creation of hazard maps by local residents in their daily life. We made an interview survey to design our system in a model traditional town in Saga, Japan. The results show that in spite of continuous efforts, many practical problems remain and residents feel unsafe. Based on these results, we designed and developed a unique ICT-based support system which contributes to community-based disaster prevention/reduction. The continuous resident participation and posting design are core concept for our sustainable community-based approach. Our system continues to support making a hazard map by integrating the community-based hazard information. Local residents register information about the spot (disaster types, a risk level, a photograph, comments, positional information) that can be dangerous in case of disaster. We have evaluated the usefulness and possibilities of our prototype system implemented as an iOS application.

The theory of partially observable Markov decision processes (POMDPs) is a useful tool for developing various intelligent agents, and learning hierarchical POMDP models is one of the key approaches for building such agents when the environments of the agents are unknown and large. To learn hierarchical models, bottom-up learning methods in which learning takes place in a layer-by-layer manner from the lowest to the highest layer are already extensively used in some research fields such as hidden Markov models and neural networks. However, little attention has been paid to bottom-up approaches for learning POMDP models. In this paper, we present a novel bottom-up learning algorithm for hierarchical POMDP models and prove that, by using this algorithm, a perfect model (i.e., a model that can perfectly predict future observations) can be learned at least in a class of deterministic POMDP environments

When developing a classifier using various objective functions, it is important to compare the performances of the classifiers. Although there are statistical analyses of objective functions for classifiers, simulation results can provide us with direct comparison results and in this case, a comparison criterion is considerably critical. A Receiver Operating Characteristics (ROC) graph is a simulation technique for comparing classifiers and selecting a better one based on a performance. In this paper, we adopt the ROC graph to compare classifiers trained by mean-squared error, cross-entropy error, classification figure of merit, and the n-th order extension of cross-entropy error functions. After the training of feed-forward neural networks using the CEDAR database, the ROC graphs are plotted to help us identify which objective function is better.

In general, computer simulations are a useful means to confirm the effectiveness of a control scheme, but the actual system is not so easy to handle. In order to overcome such underlying realistic problems, adaptability of neural networks is focused on by many researchers. In this study, an on-line training model for self-turning control is constructed based on the famous feedback error learning model, and its effectiveness is investigated through a task for controlling an inverted pendulum. As a result of several experiments, repeatable strange phenomena, such as seasonal variation of training period and reversalization of control command, are observed.

An Elman-type feedback SOM (EFSOM) is a modified version of the standard SOM for temporal signal processing. It was applied to various kinds of tasks so far, and it showed good performance on temporal signal processing. By the way, one of the major features inspired by the standard SOM is to visualize given data structure into the two-dimensional plane preserving its topology. Then, from the visualization point of view, a working mechanism of the feedback SOM (FSOM), which is an essential part of the EFSOM, is investigated in this paper. As a result of computer simulations, a trajectory of the winner neurons is bifurcated, not overlapped between any two segments, with the help of feedback pathways. This fact implies that the FSOM can deal with the past history, i.e., the context information, appropriately. Furthermore, when continuous pattern is used for training, it is found that signal processing developed in the SOM is quite similar to that for an A/D converter based on the discretized reference vectors.

In general, a biological nervous system consists of two major parts, i.e., a motor control subsystem and a sensory reception subsystem, and they work together cooperatively as direct-inverse transformation. Paying attention to this bi-directional computing manner, it is essential to perform various kinds of complex activities. Also, it must be important in most application-oriented tasks. In order to confirm its effectiveness, a bi-directional computing architecture inspired from above-mentioned scheme is constructed and applied to time series prediction tasks. At that time, one subsystem in the bi-directional model is assigned to ordinary future prediction (present→future), while the other is to past prediction (present→past) newly introduced in this study. As a result of computer simulations, it is clear that the proposed bi-directional computing architecture shows better performance than the conventional uni-directional one.

Numerous studies on time series prediction have been undertaken by a variety of researchers. Most of them typically used unidirectional computation flow, that is, present signals are applied to the model as an input and predicted future signals are derived from the model as an output. On the contrary, bidirectional computation style has been proposed recently and applied to prediction tasks. A bidirectional neural network model consists of two mutually connected subnetworks and performs direct and inverse transformations bidirectionally. To apply this model to time series prediction tasks, one subnetwork is trained a conventional future prediction task and the other is trained an additional task for past prediction. Since the coupling effects between the future and past prediction subsystems promote the model's signal processing ability, bidirectionalization of the computing architecture makes it possible to improve its performance. Furthermore, in order to investigate the acquired signal transformation, two kinds of chaotic time series—the Mackey–Glass time series and “Data Set A”—are adopted in this paper. As a result of computer simulations, it has been found experimentally that the direct and inverse transformations developed independently and their information integration give the bidirectional model an advantage over the unidirectional one. © 2003 Wiley Periodicals, Inc. Electr Eng Jpn, 145(3): 50–60, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/eej.10232

In general, a material discrimination task is achieved by following two steps: One is an analytical step for estimating constructive elements of the object, and the other is a computational step for identifying its material. The first step is simple because only measuring plural parameters is required to reach the correct answer. But the second step is complex and various kinds of techniques have been proposed. Then, one of the neural network models called a self-organizing map (SOM), a good tool for topology-preserving projection, is adopted in this paper. Firstly, each material is examined with a micro multi-functional tactile sensor. Secondly, in the training stage, the average data of several measurement trials are applied to the SOM. Finally, in the test stage, the actual measurement raw data, which are completely new to the trained SOM, are applied and the measurement object is estimated from the location of the emerging “winner” neuron. As a result of computer simulations, it is found xperimentally that measured information space is divided into several regions representing each material successfully. Furthermore, an acquired feature map for material discrimination is useful against several untraining data.

A self-organizing map (SOM) can be seen as an analytical tool to discover some underlying rules in the given data set. Based on such distinctive nature called topology-preserving projection, a new method for generating intermediate patterns was proposed. According to the results of preceding studies, most developed patterns are not morphing but dissolve. Then, in order to overcome this problem, a fragmentized distance measure is introduced in this paper. As a result of computer simulations, it is confirmed that some asymmetrical patterns are developed even though only symmetrical ones are used for training. This fact reminds us that the distance measure is quite essential, because a feature map is developed through training based on the distance measure.

An Elman-type feedback SOM (EFSOM) is a revised version of the standard SOM for dealing with time-variant information. In order to estimate its performance, an on-line character recognition task is tried with a large-scale training data set in this paper. At the same time, an idea for increasing the number of firing neurons in the competitive layer is examined to improve its trainability by means of enriching the state information. After the EFSOM is trained successfully, some validation tests for spatial displacement and temporal elasticity are also investigated. As a result of computer simulations, it is found that the EFSOM shows good performance and has good temporal signal processing ability.

A number of neural network models and training procedures for time series prediction have been proposed in the technical literature. These models studied for different time-variant data sets have typically used uni-directional computation flow or its modifications. In this study, on the contrary, the concept of bi-directional computational style is proposed and applied to prediction tasks. A bi-directional neural network model consists of two subnetworks performing two types of signal transformations bi-directionally. The networks also receive complementary signals from each other through mutual connections. The model not only deals with the conventional future prediction task, but also with the past prediction, an additional task from the viewpoint of the conventional approach. An improvement of the performance is achieved through making use of the future-past information integration. Since the coupling effects help the proposed model improve its performance, it is found that the prediction score is better than with the traditional uni-directional method. The bi-directional predicting architecture has been found to perform better than the conventional one when tested with standard benchmark sunspots data.