Masakazu MURAMATSU (村松 正和)

Offer Organization: -, System Name: -, Category: -, Fund Type: competitive_research_funding, Overall Grant Amount: - (direct: 13100000, indirect: 3930000)
錐線形計画は様々なサブクラスあるいは応用が提案されており、近年研究の発展が目覚まし
い。その中で、錐線形計画の実用化において大きな障害となっているのが「悪条件な問題」 の存在である。悪条件の極限として退化がある。悪条件あるいは退化した錐線形計画問題 は実用において頻繁に出現するにもかかわらず、従来のアルゴリズムでは解くことができ ない。
本研究は、錐線形計画問題における悪条件性に関してその理解を深め、またそのような問 題に対応する新しいアルゴリズムを開発することにより、錐線形計画の裾野を広げ、実用 に足る段階へもっていくことを目的とする。

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (C), Fund Type: -, Overall Grant Amount: - (direct: 3400000, indirect: 1020000)
Aiming to develop competitive artificial intelligence by achieving the scale-up of existing heuristics in games that are close to competition in the real world, where there are many options for actions, and situations and state transitions can only be partially observed. This research was carried out by using Go, Mahjong, a card game, chess, digital curling, etc., provided a new algorithm to realize the aim, and was published in several papers.

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (B), Fund Type: -, Overall Grant Amount: - (direct: 7200000, indirect: 2160000)
Interior-point algorithms for semidefinite programs and symmetric cone programs are analyzed in view of information geometry to show that the iteration complexity of primal-dual interior-point algorithms is approximately represented as a infomration geometric integral over central trajectory. Through extensive numerical experiments we demonstrated that the integral very accurately predict iteration-complexity of interior-point algorithms. One of the largest Gaussian graphical models in the world are successfully solved with super computer. Primal-dual interior-pont algorithms for large-scale Gaussian graphical models are developed. Regularization and facial reduction approaches for ill-conditioned semidefinite programs are developed.

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (C), Fund Type: -, Overall Grant Amount: - (direct: 3100000, indirect: 390000)
The aim of this research is to develop numerically stable primal-dual interior point methods foe solving nonlinear semidefinite programming problems. The semidefinite programming problem is an optimization problem over a closed convex cone that is not polyhedral unlike the linear programming problem. By this reason, we often observe a problem which has an asymptotic optimal solution but no optimal solution, I.e., any sequence on which the object value converges to the optimal value diverges. This brings us a numerical difficulty in determining the optimality when we apply interior point algorithms to solve the problem. A high accuracy of an optimal solution of the problem is critical if we adopt the problem as an approximation model of a combinatorial optimization problem or a robust optimization problem. To overcome the difficulty, many techniques have been proposed for obtaining a numerical stability of the algorithms. Such techniques are more highly expected when we solve nonlinear semidefinite programming problems. In order to provide one of such techniques, we introduced a homogeneous model for the nonlinear semidefinite programming problem, and showed that for the homogeneous model, (a) a bounded path having a trivial starting point exists, (b) any accumulation point of the path is a solution of the homogeneous model, c if the original problem is solvable then it gives us a finite solution, (d) if the original problem is strongly infeasible, then it gives us a finite certificate proving infeasibility, and (e) a class of algorithms for numerically tracing the path in (a) solves the problem in a polynomial number of iterations under a moderate assumption

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (C), Fund Type: -, Overall Grant Amount: - (direct: 3500000, indirect: 270000)
Markov random field (MRF) and its discriminative version have been shown useful for both biological analysis and practical applications. In biological analysis, the debate on neuronal correlations is now continuing in which the analysis of the probability P( r| s) of the neuronal response r conditional on a stimulus s is required, which could be modeled with MRF. In this context the importance of a parametric model for analyzing correlations by modeling joint probability P(r, s) is shown using Gibbs distribution.
Several approximation techniques have been proposed for computing state probabilities of MRFs, CRFs, including belief propagation, which is not applicable for MRFs in a general situation. Mean field approximation (MRF) is known as only the generally applicable approximation technique at present.
To improve the accuracy of the mean-field approximation several advanced techniques have been proposed. Since the better accuracy we attain, the more intricate equations we get into, it becomes hard to know the efficient training procedure. In fact the training procedure is known only for the naive mean-field approximation (NMF), which is not so sufficient for the approximation accuracy.
The achievement of this research is to have refined the mean field approximation to alleviate both the testing and learning time, and to have shown the efficient learning scheme for object recognition with the variational phasor mean field model (VPMF). The striking result is that our learning scheme shows comparable testing performance with SVM, despite using much smaller size of training data, and in addition the detection time and the training time are much smaller than SVM based face detection.
Performance evaluation of VPMF is given for approximation accuracy, the local minima, and a face recognition problems. We have also attained the conclusion that the correlation of population coding in neural networks is more powerful than just using only the mean firing rate.
Performance evaluation of VPMF is given for approximation accuracy, the local minima, and a face recognition problems.