Although there are straightforward methods for training neural nets to show certain types of sequential behavior, the general problem of structural and temporal credit assignment (appropriate weight changes) in networks existing in a changing environment remains unsolved. In this paper we argue that a learning mechanism for nding temporal input/output relations ought to depend solely on computations local in both space and time, and that no teacher should be required to indicate the starts and ends of relevant sequences to the network. We ask whether there are learning rules which do not depend on such external hints yet still can deal with`hidden units' and with units whose past activations arèhidden in time'. (Back propagation, for instance, is not even local in time.) We propose a discrete time version and a continuous time version of a simple on-line method local in space and time which is designed to deal with hidden states. The approach is inspired by Holland's idea of the bucket brigade for classiier systems, which is transformed to run on a neural network with xed topology. The result is a recur-rent`neural' dissipative system which is consuming`weight-substance' and permanently trying to distribute this substance onto its connections in an appropriate way. Simple experiments demonstrating the feasability of the algorithm are reported. 1 Motivation Many researchers in the eld of neural network research are concerned with the study of static input/output mappings, emphasizing the parallel aspects of such networks. Standard methods for supervised learning of such mappings are Werbos' back propagation (BP)(Werbos, 1974) and Almeida's extension (Almeida, 1987) for networks with arbitrary feedback, where the activations owing through the network have to reach a state of equilibrium. However, if we want a net to learn sequences of time-dependent outputs then an equilibrium is something we usually wish to avoid. Rumelhart, Hinton and Williams show how to correctly use BP Research supported by a scholarship of SIEMENS AG
[1]
P. Werbos,et al.
Beyond Regression : "New Tools for Prediction and Analysis in the Behavioral Sciences
,
1974
.
[2]
David Zipser,et al.
Feature Discovery by Competive Learning
,
1986,
Cogn. Sci..
[3]
John H. Holland,et al.
Properties of the Bucket Brigade
,
1985,
ICGA.
[4]
Geoffrey E. Hinton,et al.
Learning internal representations by error propagation
,
1986
.
[5]
Teuvo Kohonen,et al.
Self-Organization and Associative Memory
,
1988
.
[6]
L. B. Almeida.
A learning rule for asynchronous perceptrons with feedback in a combinatorial environment
,
1990
.