We describe a method of regularization for the restricted Bayesian network BESOM, which has network structure like deep learning. Two types of priors, win-rate penalty and lateral inhibition penalty are introduced to avoid overfitting and local minimum problems.

Multi-linear Functions (MLFs) is a well known way of probability calculation based on Bayesian Network (BN). For a givern BN, we can calculate the probability in a linear time to the size of MLF. However, the size of MLF grows exponentially with the size of BN, so the computation requires exponential time and space. Minato, et al. have shown an efficient method of calculating the probability by using Zero-Suppressed BDD (ZDD). This method is more effective than the conventional approach in some cases. In this article, we present an improvement of their method by utilizing d-separation structure of BN for efficient ZDD factorization based on weak-division operation.

The most popular algorithm for exact inference on Bayesian networks is the junction tree propagation algorithm. To improve the time and space complexity of the junction tree algorithm, we must find an optimal triangulation. For this purpose, Ottosen and Vomlel have proposed a depth-first search (DFS) algorithm for optimal triangulation using branch-and-bound and dynamic clique maintenance techniques. Nevertheless, their method entails heavy computational costs. To mitigate this problem, we propose an extended depth-first search (EDFS) algorithm. The new algorithm EDFS improves the DFS algorithm in the following two ways: (1) reduction of the computational cost of each lower bound calculation, and (2) reduction of the branching factor of each node expansion. Experimental results show that the proposed method is markedly faster than the Ottosen and Vomlel method.

Stochastic K-armed bandits tries to maximize his cumulative reward in limited number of plays. In this paper, we consider the variant of stochastic K-armed bandits that has action-dependent processing time. For this problem, we propose the policy N-UCB (Normalized UCB), the extension of well-known policy UCB, and shows some fundamental results of its regret analysis.

We propose a simpli cation method for removing irrelevant actions of a classical planning task. While previous approaches are based on reachability relations on a directed graph describing relevance of propositions and actions, our approach is based on dominance relations.

Existing grammar-based self-indexes are efficient for highly repetitive texts. However, the construction space of existing self-indexes depends on input length. Thus, developing an online construction of grammar-based self-index executed on compressed space is important for highly repetitive and streaming texts. In this paper, we present a first online grammar-based self-index named online ESP-index(OESP-index). OESP-index directly encodes an input text into a succinct representation of straight line program(SLP) in an online manner based on fully online LCA(FOLCA) techniques. The succinct representation of SLP is a wavelet tree and a bit array encoded by dynamic range min/max tree. OESP-index supports the pattern search of ESP-index by using such data structures. We experimentally show that the construction of OESP-index for real world texts is executed on compressed space.

In this paper, we are concerned with a problem of enumerating maximal k-plexes in a given undirected graph, where the notion of k-plex is a relaxation model of clique. An existing algorithm for this task tries to nd solution k-plexes based on a theoretical property of diameter of k-plex . In order to enjoy this propery fully, we propose to divide the class of maximal k-plexes into several subclasses based on size and properness of k-plexes . We, then, observe the maximal k-plexes in each subclass have smaller diameter, where the smaller diameter of k-plex becomes, the less the number of search branches becomes. As a result, we expect that computational cost for our enumeration task can be reduced. Our experimental results for several benchmark graphs show that the proposed approach can achieve a certain degree of improvement in efficiency.

Many variants of pseudo-cliques have been introduced as a relaxation model of cliques to detect communities in real world networks. For most types of pseudo-cliques, enumeration algorithms can be designed just similar to maximal clique enumerator. However, the problem of enumerating pseudo-cliques is computational hard, because the number of maximal pseudo-cliques-cliques is huge in general. Furthermore, because of the weak requirement of k-plex, sparse communities are also allowed depending on the parameter k. To obtain a class of more dense pseudo cliques and to improve the computational performance, we introdue a derived graph whose vertices are cliques in the original input graph. Then our target pseudo must be a clique or a pseudo clique of the derived graph under an additional constraint requiring density in the original graph. An enumerator for this new class is designed and its computational efficientcy is experimentally verfied.