A probabilistic automaton with pattern recognition supervision is considered as an on-line real-time adaptive controller for a complex plant with a multimodal performance-index structure and subjected to an environment which randomly fluctuates in time. This environment is considered to be partially measurable but entirely uncontrollable. The automaton discussed is capable not only of learning the optimum control parameters in any given environmental situation but also of acting as an internal teacher in the formation of pattern associations between the measurable state of the environment and the control situation, so that approximately recurrent conditions can be taken advantage of in future relearning situations. These pattern associations, once developed, are used to supervise the future action of the automaton. Furthermore, the pattern associations between the measurable state of the environment and the control situation must themselves be adaptively formed to allow for variations caused by unknown and/or unmeasurable factors in the total environment.
[1]
Samuel H. Brooks.
A Discussion of Random Methods for Seeking Maxima
,
1958
.
[2]
George S. Sebestyen,et al.
Pattern recognition by an adaptive process of sample set construction
,
1962,
IRE Trans. Inf. Theory.
[3]
J. L. Douce,et al.
The development and performance of a self-optimizing system
,
1963
.
[4]
Dean C. Karnopp,et al.
Random search techniques for optimization problems
,
1963,
at - Automatisierungstechnik.
[5]
Harold J. Kushner,et al.
A New Method of Locating the Maximum Point of an Arbitrary Multipeak Curve in the Presence of Noise
,
1964
.
[6]
J. C. Hill,et al.
Hillclimbing on Hills with Many Minima
,
1965
.
[7]
King-Sun Fu,et al.
A variable structure automaton used as a multi-modal searching technique
,
1965
.