In this paper we will analyze how supervised learning occurs in ecological neural networks (i.e. networks that interact with an external environment). Using an evolutionary method for selecting good teaching inputs we will show how the learning process interacts with the capability of such networks to partially determine the next input stimuli with their outputs. In trying to explain the behavior of these networks we surprisingly find that for obtaining a desired output X it is better to use a teaching input different from X. To explain this fact we claim that teaching inputs in ecological networks have two different effects; (a) to reduce the discrepancy between the actual output of the network and the teaching inputs themselves, (b) to modify the network behavior and as a consequence the network learning experiences. Evolved teaching inputs appear to represent a compromise between these two needs. We finally show how evolved teaching inputs that are allowed to change during the learning process respond differently in different period of learning first giving more weight to the (b) function and progressively later on to the (a) function. The notion of ecological neural networks refers to an approach to the study of neural networks that views network as behaving, learning, developing and evolving in an environment (see Parisi, Cecconi and Nolfi, 1990). Hence, within this framework, the behavior of a network tends to be studied with reference to the environment in which the network behaves. The most important consequence of behaving in an environment is that the output of an ecological network partially determines the network's input. By acting on the environment with its motor output an ecological network may change the environment (i.e. the network can modify the position or the characteristics of an object in the environment) or it may change its relation to the environment (i.e. by moving itself the network can modify the angle of an object with respect to the direction it faces or even move to a different environment). Therefore sensory input becomes a function of the independent properties of the environment and the network's behavior. food element. We shall also equip O with a simple motor system that provides it with the possibility, in a single action, to turn any angle from 90 degrees left to 90 degrees right and then move from 0 to 5 cells forward. Finally, when O happens to step on a food cell, it eats the food element which disappears. We imagine that a network represents the nervous system of an organism (O) and that O's environment is a two-dimensional square divided up into cells. At any particular moment O occupies one of these cells. A number of food elements are randomly distributed in the environment with each food element occupying a single cell. O has a facing direction. We shall imagine it has a rudimentary sensory system that allows it to receive as input from the environment the angle (relative to where O is currently facing) and the distance of the nearest Figure 1. Auto-teaching network The network underlying O's behavior is a feedforward network consisting of three layers (Figure 1). The input layer includes 2 units which receive sensory information from the environment. These 2 units are fully connected with two sets of intermediate ("hidden") layers of 7 units. The first set of hidden units is connected with 2 output units that code O's movement.
[1]
Geoffrey E. Hinton,et al.
Learning internal representations by error propagation
,
1986
.
[2]
Stefano Nolfi,et al.
Econets: Neural networks that learn in an environment
,
1990
.
[3]
John Maynard Smith,et al.
When learning guides evolution
,
1987,
Nature.
[4]
John H. Holland,et al.
Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence
,
1992
.
[5]
V. Marchman,et al.
U-shaped learning and frequency effects in a multi-layered perception: Implications for child language acquisition
,
1991,
Cognition.
[6]
Richard K. Belew,et al.
Evolving networks: using the genetic algorithm with connectionist learning
,
1990
.
[7]
David H. Ackley,et al.
Interactions between learning and evolution
,
1991
.