This paper proposes the autocorrelation function (acf) and partial autocorrelation function (pacf) as tools to help and improve the construction of the input layer for univariate time series artificial neural network (ANN) models, as used in classical time series analysis. Especially reducing the number of input layer neurons, and also helping the user to understand the behaviour of the series. Although the acf and pacf are considered linear functions, this paper shows that they can be used even in non linear time series. The ANNs used in this work are the Incremental Gaussian Mixture Network (IGMN), because it is a deterministic model, and the multilayer perceptron (MLP), the most used ANN model for time series forecasting.

In this work we use IGMN (standing for incremental Gaussian mixture network), an incremental neural network model based on Gaussian mixtures, for on-line control and robotics. IGMN is inspired on recent theories about the brain, specially the memory-prediction framework and the constructivist artificial intelligence, which endows it with some unique features that are not present in most artificial neural network models. Moreover, IGMN learns incrementally from data flows (each data can be immediately used and discarded) and asymptotically converges to the optimal regression surface as more training data arrive. Through several experiments using the proposed model in robotics it is demonstrated that IGMN is not sensitive to initialization conditions, does not require fine-tuning its configuration parameters and has a good computational performance, thus allowing its use in real time control applications.

In this work, a one-shot learning solution to the t-maze road sign problem is presented. This problem consists in taking the correct turning decision at a bifurcation after seeing a light signal some time steps before. The recently proposed Echo State Incremental Gaussian Mixture Network (ESIGMN) is used in order to learn the correct behavior after a single scan through a single training example (and its mirrored version) generated by a simple reactive controller. Experiments with different time delays between the signal and the decision point are performed, and the ESIGMN is shown to solve the problem while achieving good performance. This one-shot ability can be useful for online learning in robotics, since the robot can learn with minimum interaction with the environment.

Boundary Value Problem (BVP) Path Planners generate potential fields whose gradient descent represents navigational routes from any point of the environment to a goal position. The resulting trajectories are smooth and free of local minima. A BVP planner has been recently used to compute trajectories in known inhomogeneous environments, corresponding to different degrees of traveling difficulties. In this case, we locally distort the potential field generating regions with high or low preferences for navigating. In this paper, we extend this idea, generating an strategy allowing the robot to explore an unknown environment dynamically considering environment preferences. Several experiments demonstrate that our method can be used as a core in an integrated exploration approach.

The large amount of patterns generated by frequent pattern mining algorithms has been extensively addressed in the last few years. In geographic pattern mining, besides the large amount of patterns, many are well known geographic domain associations. Existing algorithms do not warrant the elimination of all well known geographic dependences since no prior knowledge is used for this purpose. This paper presents a two step method for mining frequent geographic patterns without associations that are previously known as non-interesting. In the first step the input space is reduced as much as possible. This is as far as we know still the most efficient method to reduce frequent patterns. In the second step, all remaining geographic dependences that can only be eliminated during the frequent set generation are removed in an efficient way. Experiments show an elimination of more than 50% of the total number of frequent patterns, and which are exactly the less interesting.

In frequent geographic pattern mining a large amount of patterns is well known a priori. This paper presents a novel approach for mining frequent geographic patterns without associations that are previously known as non- interesting. Geographic dependences are eliminated during the frequent set generation using prior knowledge. After the dependence elimination maximal generalized frequent sets are computed to remove redundant frequent sets. Experimental results show a significant reduction of both the number of frequent sets and the computational time for mining maximal frequent geographic patterns.

In this paper we introduce RL-CD, a method for solving reinforcement learning problems in non-stationary environments. The method is based on a mechanism for creating, updating and selecting one among several partial models of the environment. The partial models are incrementally built according to the system's capability of making predictions regarding a given sequence of observations. We propose, formalize and show the efficiency of this method both in a simple non-stationary environment and in a noisy scenario. We show that RL-CD performs better than two standard reinforcement learning algorithms and that it has advantages over methods specifically designed to cope with non-stationarity. Finally, we present known limitations of the method and future works.

We propose some techniques to improve the exploration method based on BVP and demonstrate that, at least, for exploration in two dimensional space, we can reduce enormously the computation cost of the harmonic function method for exploration without losing the important properties of the method. Switching between local and global potential calculation and using an adaptive activation window result in a computationally efficient and refined mapping process.

The use of relaxation methods for calculation of harmonic potentials has proved to be a powerful technique for path planning in a known environment and recently for exploration of unknown environments. In the latter case, the potential is calculated on partial versions of the map represented on an occupancy grid, and it indicates safe paths towards the unexplored regions. Here, we show that exploration based on potentials functions calculated from boundary value problems works fairly well in sparse environments provided that we no longer use harmonic potentials. We propose different families of potentials that improve the exploration process in this kind of environment, while keeping the important property of not having dynamical local minima.

The use of relaxation methods for calculation of harmonic potentials has proved to be a powerful technique for path planning in a known environment. We show that this idea can be successfully extended to exploration of unknown environments. The potential is calculated in a partial version of the map, represented on an occupancy grid, and it indicates safe paths towards the unexplored regions. We demonstrate that a complete relaxation of the potential is not necessary to accomplish smooth performances. Furthermore, we discuss the effect of different relaxation methods in the calculation of harmonic potential.