Least-squares policy iteration is a useful reinforcement learning method in robotics due to its computational efficiency. However, it tends to be sensitive to outliers in observed rewards. In this paper, we propose an alternative method that employs the absolute loss for enhancing robustness and reliability. The proposed method is formulated as a linear programming problem which can be solved efficiently by standard optimization software, so the computational advantage is not sacrificed for gaining robustness and reliability. We demonstrate the usefulness of the proposed approach through a simulated robot-control task.

A robust routing algorithm was developed based on reinforcement learning that uses (1) reward-weighted principal component analysis, which compresses the state space of a network with a large number of nodes and eliminates the adverse effects of various types of attacks or disturbance noises, (2) activity-oriented index allocation, which adaptively constructs a basis that is used for approximating routing probabilities, and (3) newly developed space compression based on a potential model that reduces the space for routing probabilities. This algorithm takes all the network states into account and reduces the adverse effects of disturbance noises. The algorithm thus works well, and the frequencies of causing routing loops and falling to a local optimum are reduced even if the routing information is disturbed.

An orthonormal basis adaptation method for function approximation was developed and applied to reinforcement learning with multi-dimensional continuous state space. First, a basis used for linear function approximation of a control function is set to an orthonormal basis. Next, basis elements with small activities are replaced with other candidate elements as learning progresses. As this replacement is repeated, the number of basis elements with large activities increases. Example chaos control problems for multiple logistic maps were solved, demonstrating that the method for adapting an orthonormal basis can modify a basis while holding the orthonormality in accordance with changes in the environment to improve the performance of reinforcement learning and to eliminate the adverse effects of redundant noisy states.

In this paper, an adaptive bound reduced-form genetic algorithm (ABRGA) to tune the control points of B-spline neural networks is proposed. It is developed not only to search for the optimal control points but also to adaptively tune the bounds of the control points of the B-spline neural networks by enlarging the search space of the control points. To improve the searching speed of the reduced-form genetic algorithm (RGA), the ABRGA is derived, in which better bounds of control points of B-spline neural networks are determined and paralleled with the optimal control points searched. It is shown that better efficiency is obtained if the bounds of control points are adjusted properly for the RGA-based B-spline neural networks.

This paper proposes a simple learning algorithm that can realize any boolean function using the three-layer binary neural networks. The algorithm has flexible learning functions. 1) moving "core" for the inputs separations,2) "don't care" settings of the separated inputs. The "don't care" inputs do not affect the successive separations. Performing numerical simulations on some typical examples, we have verified that our algorithm can give less number of hidden layer neurons than those by conventional ones.

A function approximation scheme for image restoration is presented to resolve conflicting demands for smoothing within each object and differentiation between objects. Images are defined by probability distributions in the augmented functional space composed of image values and image planes. According to the fuzzy Hough transform, the probability distribution is assumed to take a robust form and its local maxima are extracted to yield restored images. This statistical scheme is implemented by a feedforward neural network composed of radial basis function neurons and a local winner-takes-all subnetwork.

In previous papers the building of hierarchical networks made up of components using fuzzy rules was presented. It was demonstrated that this approach could be used to construct networks to solve classification problems, and that in many cases these networks were computationally less expensive and performed at least as well as existing approaches based on feedforward neural networks. It has also been demonstrated how this approach could be extended to real-valued problems, such as function approximation and time series prediction. This paper investigates the problem of choosing the best network for real-valued approximation problems. Firstly, the nature of the network parameters, how they are interrelated, and how they affect the performance of the system are clarified. Then we address the problem of choosing the best values of these parameters. We present two model selection tools in this regard, the first using a simple statistical model of the network, and the second using structural information about the network components. The resulting network selection methods are demonstrated and their performance tested on several benchmark and applied problems. The conclusions look at future research issues for further improving the performance of the clustering network.