A list of my intelligent control publications, followed by their citations and abstracts.
Lately, there has been an extensive interest in the possible uses of neural networks for nonlinear system identification and control. In this paper, we provide a framework for the simultaneous identification and control of a class of unknown, uncertain nonlinear systems. The identification portion relies on modeling the system by a neural network which is trained via a local variant of the Extended Kalman Filter. We will discuss this local algorithm for training a neural network to approximate a nonlinear feedback system. We also give a dynamic programming-based method of deriving near optimal control inputs for the real plant based on this approximation and a measure of its error (covariance). Finally, we combine these methods in a hierarchical algorithm for identification and control of a class of uncertain, unknown systems. The complexity of the whole algorithm is analyzed.
In this paper I describe some results obtained from experiments with task-level learning [C.G. Atkeson, E.W. Aboaf, S.M. Drucker]. The main idea of task-level learning is that a given task, e.g., throwing a ball into a basket, can be viewed as an input/output system driven by a vector of input variables or commands, c, and responding with a vector of output variables or performance indicators, p. This formulation allows us to bring powerful numerical methods to bear on problems at a high-level of performance measurement: the task level. We study task-level learning as a paradigm that may help us to program machines to learn from experience: (1) perform a task better over time, (2) optimize task performance, and (3) generalize knowledge over tasks.
Some extensions to the paradigm are explored in this paper: (1) A new refined model learning scheme is presented. Simulation experiments are performed to test the effects of different inverse models, different learning shemes, and different learning intervals. (2) A framework for dealing with tasks that inherently try to minimize or maximize performance is presented.