Learned Evaluation and Sequential Credit Assignment
The model, as developed so far, is capable of learning arbitrary S-R mappings, provided that it is explicitly taught the correct function. The need for such an omniscient instructor is a common criticism of neural learning “with a teacher”. However, the occasional argument that learning is unnecessary if the correct response is already known, ignores the fact that instruction may occur only after a response is made. The correct response is not known beforehand, but may be at least partially deduced from succeeding state changes. This is the essence of trial-and-error learning, a basic feature of biological behavior. Post-response instruction may completely determine the correct output (as in Aplysia gill withdrawal) or may only evaluate the correctness of the preceding action (sometimes referred to as “learning with a critic” (Widrow et al., 1973; Barto et al., 1981)). A specialized evaluation system is developed in this chapter and used to train the operators.
KeywordsConditional Stimulus Boolean Function Learned Evaluation Shared Memory Classical Conditioning
Unable to display preview. Download preview PDF.