OPERANT-CONDITIONABLE MACHINES.
Abstract
The term 'operant-conditionable machine' refers to an adaptive controller in which adaptation proceeds by the application of generalized 'punishments' or 'rewards.' The machine is based on a network of 'neuromimes,' electronic circuits which mimic the function of living neurons. Each neuromime can act upon others in either an excitatory or an inhibitory manner. Adaptation or training takes place by adjustment of thresholds of the individual neuromimes. The punishment or reward is connected to all neuromimes in the net, but only those which are currently active are able to accept the reinforcement signal. In this way, the subpopulation of neuromimes which responds to a given input can be induced to respond more or less strongly. In the latter case (punishment), the reduction in strength of response will reduce the inhibitory influence the subpopulation was exerting on the rest of the network. A new subpopulation will then respond; therefore, the net will have been induced to try a new response to the same input. A breadboard conditionable machine of limited size has been built and has been used to demonstrate a number of trainable functions. Any one of a number of inputs can be selectively connected to a single output by application of training signals. Any one of a number of outputs can be connected to a single input. Temporal waveforms can be generated and can be modified by application of reinforcements. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1970
- Accession Number
- AD0703758
Entities
People
- Charles E. Hendrix