Nowadays, deep learning and reinforcement learning have given high performance in various fields, and attracted much attention. In order to apply these learning methods to real problems, they must have a sufficient generalization ability. Whereas to improve the generalization ability has been actively studied in some fields such as image recognition and speech recognition, it has not been sufficiently studied for sequential decision-making problems such as game play and path finding. This paper proposes a reinforcement learning method with the generalization ability developed by using deep learning for a path finding problem, which is one of the sequential decision-making problems. Experimental results show that the generalization ability of the proposed method is superior to that of deep learning methods and deep reinforcement learning methods.

The simulations of complex social systems involving diverse stakeholders can be evaluated from multiple analytical interests. In this paper, we propose a method to classify social simulation logs hierarchically along with multiple analytical interests and visualize them based on the frequency of classification results. The proposed method introduces the concept of cladistic phylogeny and classifies social simulation logs hierarchically, in such a way as to classify species. In the proposed method, analysts make a hierarchical classification diagram of social simulation logs called a possibility cladogram and understand the frequency of possible results from the viewpoint of analytical interests intuitively. We applied the proposed method to Schelling's segregation model. We confirmed that the proposed method helps intuitively understand the frequency of possible results from the viewpoint of analytical interests and also give assistance to run efficient microdynamics analysis by referring to the classification result of simulation logs on the possibility cladogram. We evaluated the proposed method from two perspectives: realization of visualization benefits corresponding to usefulness and validity as a visualization method.

Bounded real lemma is discussed for interval delay systems, and a necessary and sufficient condition is characterized with parameter-dependent LMIs. The condition is further transformed to standard LMIs, and it is shown that the parameter-region partitioning enables to treat the necessary and sufficient condition with arbitrary resolution and precision. The computation efficiency for the evaluation of optimal H_∞ performance is discussed with numerical examples.

In biological neural networks of living organisms, various firing patterns of nerve cells have been observed, typical examples of which are burst firings and periodic firings. In this paper we propose a learning method which can realize various firing patterns for recurrent spiking neural networks (RSNNs). We have already proposed learning methods of RSNNs in which the learning problem is formulated such that the number of spikes emitted by a neuron and their firing instants coincide with given desired ones. In this paper, in addition to that, we consider several desired properties of a target RSNN and propose cost functions for realizing them. Since the proposed cost functions are not differentiable with respect to the learning parameters, we propose a learning method based on the particle swarm optimization. Furthermore we apply the proposed method to developing a model for “visual feature extraction” in biological system and demonstrate its effectiveness.