J. A. Becerra

Multimodule Artificial Neural Network Architectures for Autonomous Robot Control Through Behavior Modulation

In this paper we consider one of the big challenges when constructing modular behavior architectures for the control of real systems, that is, how to decide which module or combination of modules takes control of the actuators in order to implement the behavior the robot must perform when confronted with a perceptual situation. The problem is addressed from the perspective of combinations of ANNs, each implementing a behavior, that interact through the modulation of their outputs. This approach is demonstrated using a three way predator-prey-food problem where the behavior of the individual should change depending on its energetic situation. The behavior architecture is incrementally evolved.

Towards real-time hyperspectral image processing, a GP-GPU implementation of target identification

In the quest for real time processing of hyperspectral images, this paper presents two artificial intelligence algorithms for target detection specially developed for their implementation over GPU and applied to a search-and-rescue scenario. Both algorithms are based on the application of artificial neural networks to the hyperspectral data. In the first algorithm the neural networks are applied at the level of individual pixels of the image. The second algorithm is a multiresolution based approach to scale invariant target identification using a hierarchical artificial neural network architecture. We have studied the main issues for the efficient implementation of the algorithms in GPU: the exploitation of thousands of threads that are available in this architecture and the adequate use of bandwidth of the device. The tests we have performed show both the effectiveness of detection of the algorithms and the efficiency of the GPU implementation in terms of execution times.

Induced behavior in a real agent using the multilevel darwinist brain

In this paper we present a strategy for inducing a behavior in a real agent through a learning process with a human teacher. The agent creates internal models extracting information from the consequences of the actions it must carry out, and not just learning the task itself. The mechanism that permits this background learning process is the Multilevel Darwinist Brain, a cognitive mechanism that allows an autonomous agent to decide the actions it must apply in its environment in order to fulfill its motivations. It is a reinforcement based mechanism that uses evolutionary techniques to perform the on line learning of the models.

Complex behaviours through modulation in autonomous robot control

Combining previous experience and knowledge to contemplate tasks of increasing complexity is one of the most interesting problems in autonomous robotics. Here we present an ANN based modular architecture that uses the concept of modulation to increase the possibilities of reusing previously obtained modules. A first approximation to the modulation of the actuators was tested in a previous paper where we showed how it was useful to obtain more complex behaviours that those obtained using only activation / inhibition. In this paper we extend the concept to sensor modulation, which enables the architecture to easily modify the required behaviour for a module, we show how both types of modulation can be used at the same time and how the activation / inhibition can be seen as a particular case of modulation. Some examples in a real robot illustrate the capabilities of the whole architecture.

Neural clustering analysis of macroevolutionary and genetic algorithms in the evolution of robot controllers

In this work, we will use self-organizing feature maps as a method of visualization the sampling of the fitness space considered by the populations of two evolutionary methods, genetic and macroevolutionary algorithms, in a case with a mostly flat fitness landscape and low populations. Macroevolutionary algorithms will allow obtaining better results due to the way in which they handle the exploration-exploitation equilibrium. We test it with different alternatives using the self-organizing maps.

A profiling based intelligent resource allocation system

The work presented here is mainly concerned with the development of an intelligent resource allocation method specially focused in providing maximum satisfaction to user agents tied to resource strapped applications. One of the applications of this type of strategies is that of remote sensing in terms of energy and sensor usage. Many remote sensors or sensor arrays reside on satellites and their use must be economized, while at the same time the agency managing satellite time would like to satisfy the users as much as possible. Here we have developed a cognitive based strategy that obtains models of users and resource use in real time and uses these models to obtain strategies that are compatible with management policies. The paper concentrates in obtaining the user models.

Evolution of cooperating ANNs through functional phenotypic affinity

This work deals with the problem of automatically obtaining ANNs that cooperate in modelling of complex functions. We propose an algorithm where the combination of networks takes place at the phenotypic operational level. Thus, we evolve a population of networks that are automatically classified into different species depending on the performance of their phenotype, and individuals of each species cooperate forming a group to obtain a complex output. The components that make up the groups are basic ANNs (primitives) and could be reused in other search processes as seeds or could be combined to generate new solutions. The magnitude that reflects the difference between ANNs is their affinity vector, which must be automatically created and modified. The main objective of this approach is to model complex functions such as environment models in robotics or multidimensional signals.

Some approaches for reusing behaviour based robot cognitive architectures obtained through evolution

This chapter provides a vision of some of the work we have been carrying out with the objective of making evolutionarily obtained behaviour based architectures and modules for autonomous robots more standardized and interchangeable. These architectures are based on a hierarchical behaviour structure where all of the modules, as well as their interconnections, are automatically obtained through evolutionary processes. The objective has been to obtain practical structures that would work in real robots operating in real environments and it is a first step towards a more ambitious approach in which no inkling to which would be the optimal organization of the modules would be provided. The emphasis of this work is to produce behaviour based structures that work on real robots operating in real environments and to be able to obtain them as independent of the platform as possible. To address this problem we have introduced the concept of virtual sensors and effectors in behaviour based architectures and studied different approaches to automatically obtain them.