The modeling of nonlinear dynamical systems is one of the emergent application areas of artificial neural networks. In this paper, we present a general methodology based on neural networks and genetic algorithms that can be applied to modeling of nonlinear dynamical systems. We describe a general methodology for modeling nonlinear systems with known rank (i.e. state-space dimension) by feedforward networks with external delay units. We point out the shortcomings of this approach when the rank of the system is not known a priori. In this case, it is beneficial to employ genetic algorithms to search for neural networks that can model the nonlinear dynamical systems. Two genetic algorithms are presented for this case: one that determines the best feedforward network with external delay, and one that searches for a network with arbitrary topology and memory cells within each neuron.

Genetic algorithm based neural networks for dynamical system modeling

This paper gives the concept of an autonomous robotic agent that is capable of showing both machine learning and reactive behavior. The first methodology is used to collect information about the environment and to plan robot actions based on this information while the robot is performing tasks. Processing and storing information obtained during several task executions is called lifelong learning. Reactive behavior, the second desirable feature of an autonomous robot, is needed to execute the actions in a dynamically changing environment. This paper presents the main components of the machine learning part of the autonomous robotic agent: the conceptual state generalization module that forms conceptual states as categories (clusters) of observation vectors obtained from the sensor system; and the robot behavior effects modeller that calculates the next conceptual state from the current state and an action obtained from the action planner.

Neural network-based modeling of robot action effects in conceptual state space of real world

We present the concept of an intelligent robotic agent that displays both learning-based and reactive capabilities. The learning-based components of the agent enable it to build up action strategies over its lifetime in a real world environment. Due to its complexity and dynamically changing nature, it is impossible to gather full knowledge about the environment. This necessitates the addition of reactive components that enable the robotic agent to carry out tasks in the presence of environment uncertainties. Our view of an intelligent robotic agent is based on the decomposition into a learning-based action planner and a reactive action executor part. We present an overview of the modules that implement the functionality of these two main components.

Intelligent robotic agent combining reactive and cognitive capabilities

This paper presents the concept of an autonomous robotic agent combining reactive and machine learning-based algorithms. The focus is on the machine learning-based part that we implement by neural networks. A method for reducing the environment state space to a smaller conceptual world space is given. We then show how the concept of "lifelong learning" can be implemented by neural networks in a robotic action planner.

Robotic agent control combining reactive and learning capabilities

With the development of new computing paradigms, such as neural networks and genetic algorithms, new tools have become available in computer-aided systems theory. These tools can be used to tackle problems that are considered “hard” in traditional systems theory, like the modeling and identification of nonlinear dynamical systems.

Stephan Dreiseitl

Papers

Hybrid Evolutionary Programming: the Tools for CAST