ABSTRACT
A good deal of speculation has gone into the subject of mechanizing learning processes. The various schemes known to the author may be shown to be mechanizations of some general plan devised by the designer for analyzing the environment with which the device is to deal. The method proposed in this paper is another such plan. It is hoped, however, that the rules for doing this are simpler than most and that the environment in which the device is designed to operate has some small trace of generality.
Let it be required to produce a device which can be used to analyze an environment which consists of a finite number of variables, each of which may at any instant assume one of a finite number of conditions. It is not necessary that each of the variables be able to assume the same number of distinct conditions, nor is it required that all or any fixed number of the variables be present at any time. It is also assumed that the changes in the variables are made instantaneously and if more than one variable changes, all may change instantaneously together if so desired. Let the analyzing device be able to observe the various changes that take place in the environment. Also let it be required that the learning device be able to find some functional relationships between the variables such that if all but one of the variables are known, the condition of that one may be found from a knowledge of the other variables. Three cases are apparent. 1. The environment is entirely predictable on observation of a given state. 2. The environment is partially predictable. 3. The environment is completely random in action and thus is unpredictable. Case 3 is of no interest because no relationships exist which are inherent in the environment. Case 2 may be usefully examined in an attempt to note the degree of predictability present. However, case 1 will be treated first as it may be used to demonstrate the learning process to its greatest advantage.
Index Terms
- A learning process suitable for mechanization
Recommendations
Level-dependent Sugeno integral
In this paper, a new concept of level-dependent Sugeno integral is introduced, and it is used to represent comonotone maxitive aggregation functions acting on a complete scale K. The relationship between the level-dependent Sugeno integral and some ...
Indicators of `innovation as a process'
Innovation as a process is related to the business viewed as a process. A process cannot be captured through the indicators of input/output, which are the most commonly accepted variables. Indicators of technological characteristics also limit the scope ...
Comments