In this paper, we present an algorithm for learning non-linear dynamical systems which works by aligning local linear models, based on a probabilistic formulation of subspace identification. This is achieved by the fusion of the recent works in the fields of machine learning and system control. Because the procedure for constructing a state sequence in subspace identification can be interpreted as the Canonical Correlation Analysis(CCA) between past and future observation sequences, we can derive a latent variable representation for this problem. Therefore, as in a similar manner to the recent works on learning a mixture of probabilistic models, we obtain a framework for constructing a state space by aligning local linear coordinates. This leads to a prominent algorithm for learning non-linear dynamical systems. Finally, we apply our method to motion capture data and telemetry data, and then show how our algorithm works well.

View full abstract

Relational similarity can be defined as the similarity between two semantic relations R and R' that exist respectively in two word pairs (A,B) and (C,D). Relational search, a novel search paradigm that is based on the relational similarity between word pairs, attempts to find a word D for the slot ? in the query {(A,B), (C,?)} such that the relational similarity between the two word pairs (A, B) and (C, D) is a maximum. However, one problem frequently encountered by a Web-based relational search engine is that the inherent noise in Web text leads to incorrect measurement of relational similarity. To overcome this problem, we propose a method for verifying a relational search result that exploits the symmetric properties in proportional analogies. To verify a candidate result D for a query {(A, B), (C, ?)}, we replace the original question mark by D to create a new query {(A,B),(?,D)} and verify that we can retrieve C as a candidate for the new query. The score of C in the new query can be seen as a complementary score of D because it reflects the reliability of D in the original query. Moreover, transformations of words in proportional analogies lead to relational symmetries that can be utilized to accurately measure the relational similarity between two semantic relations. For example, if the two word pairs (A,B) and (C, D) show a high degree of relational similarity then the two word pairs (B,A) and (D,C) also have a high degree of relational similarity. We apply this idea in relational search by using symmetric queries such as {(B, A), (D, ?)} to create six queries for verifying a candidate answer D to improve the reliability of the verification process. Our experimental results on the Scholastic Aptitude Test (SAT) analogy benchmark show that the proposed method improves the accuracy of a relational search engine by a wide margin.

View full abstract

Distributed sensor network is an important research area of multi-agent systems. We focus on a type of distributed sensor network systems that cooperatively observe multiple objects using multiple autonomous gaze control sensors. The sensor resource allocation problems of the distributed sensor network can be formalized as distributed constraint optimization problems. However, in the previous works, computation cost to solve the resource allocation problems highly increases with scale/density of the problems. In this work, we divide the problem into two layers of problems. Then two layered cooperative solvers are applied to those problems. Moreover, constraints to keep stability of the allocation in dynamic environment are introduced. The result of the experiment shows that proposed method reduces the number of message cycles. Effects of constraints for the stability of the allocation are also shown.

View full abstract

This paper describes a computational customer behavior modeling by Bayesian network with an appropriate category. Categories are generated by a heterogeneous data fusion using an ID-POS data and customer's questionnaire responses with respect to their lifestyle. We propose a latent class model that is an extension of PLSI model. In the proposed model, customers and items are classified probabilistically into some latent lifestyle categories and latent item category. We show that the performance of the proposed model is superior to that of the k-means and PLSI in terms of category mining. We produce a Bayesian network model including the customer and item categories, situations and conditions of purchases. Based on that network structure, we can systematically identify useful knowledge for use in sustainable services. In the retail service, knowledge management with point of sales data mining is integral to maintaining and improving productivity. This method provides useful knowledge based on the ID-POS data for efficient customer relationship management and can be applicable for other service industries. This method is applicable for marketing support, service modeling, and decision making in various business fields, including retail services.

View full abstract

We address a dynamic decision problem in which decision makers must pay some costs when they change their decisions along the way. We formalize this problem as Dynamic SAT (DynSAT) with decision change costs, whose goal is to find a sequence of models that minimize the aggregation of the costs for changing variables. We provide two solutions to solve a specific case of this problem. The first uses a Weighted Partial MaxSAT solver after we encode the entire problem as a Weighted Partial MaxSAT problem. The second solution, which we believe is novel, uses the Lagrangian decomposition technique that divides the entire problem into sub-problems, each of which can be separately solved by an exact Weighted Partial MaxSAT solver, and produces both lower and upper bounds on the optimal in an anytime manner. To compare the performance of these solvers, we experimented on the random problem and the target tracking problem. The experimental results show that a solver based on Lagrangian decomposition performs better for the random problem and competitively for the target tracking problem.

View full abstract