Search

We present a method of nominally piecewise multiple regression using a four-layer perceptron to fit multivariate data containing numerical and nominal variables. In our method, each linear regression function is accompanied with the corresponding nominal condition stating a subspace where the function is applied. Our method selects the optimal numbers of hidden units and rules very fast based on the Bayesian Information Criterion (BIC). The proposed method worked well in our experiments using an artificial and two real data sets. [ABSTRACT FROM AUTHOR]

We present a method for discovering nominally conditioned polynomials to fit multivariate data containing numeric and nominal variables using a four-layer perceptron having shared weights. A polynomial is accompanied with the nominal condition stating a subspace where the polynomial is applied. To get a succinct neural network, we focus on weight sharing, where a weight is allowed to have one of common weights. A near-zero common weight can be eliminated. Our method iteratively merges and splits common weights based on 2nd-order criteria, escaping from local optima. Moreover, our method selects the optimal number of hidden units based on cross-validation. The experiments showed that our method can restore the original sharing structure for an artificial data set, and discovers rather succinct rules for a real data set. [ABSTRACT FROM AUTHOR]

We have proposed an online policy-improving system of reinforcement learning (RL) agents with a mixture model of Bayesian Networks (BNs), and discussed properties of the system. In this paper, two types of mixture models have been applied to the system. A structure of BN in the mixture model is selected based on data collected by agents in an environment, and is regarded as a stochastic knowledge of the environment. This research investigates the adaptability of our system to dynamic environments containing an unexperienced environment, in which an agent does not have the knowledge. [ABSTRACT FROM AUTHOR]