|
|
 |
|
|
|
|
4. The performance measure varies over time and space so that given adaptations are only advantageous at certain places and times. |
|
|
|
|
|
|
|
|
5. The environment E presents to t a great flux of information (including performances) which must be filtered and sorted for relevance. |
|
|
|
|
|
|
|
|
By describing these obstacles, and the adaptive plans meant to overcome them, within a general framework, we open the possibility of discovering plans useful in any situation requiring adaptation. |
|
|
|
|
|
|
|
|
Before going further let us flesh out these abstractions by using them in the description of two distinct adaptive systems, one simple and artificial, the other complex and natural. |
|
|
|
|
|
|
|
|
3. A Simple Artificial Adaptive System |
|
|
|
|
|
|
|
|
The artificial adaptive system of this example is a pattern recognition device. (The device to be described has very limited capabilities; while this is important in applications, it does not detract from the device's usefulness as an illustration.) The information to be fed to the adaptive device is preprocessed by a rectangular array of sensors, a units high by b units wide. Each sensor is a threshold device which is activated when the light falling upon it exceeds a fixed threshold. Thus, when a "scene" is presented to the sensor array at some time t,each individual sensor is either "on" or "off" depending upon the amount of light reaching it. Let the activity of the ith sensor, i = 1, 2, . . ., ab,at time t be represented formally by the function di(t), where di(t)= 1 if the sensor is "on" and di(t)= 0 if it is ''off." A given scene thus gives rise to a configuration of ab "ones" and "zeros." All told there are 2ab possible configurations of sensor activation; let C designate this set of possible configurations. It will be assumed that a particular subset of C1of C corresponds to (instances of) the pattern to be recognized. The particular subset involved, among the 22ab possible, will be unknown to the adaptive device. (E.g., C1 might consist of all configurations containing a connected X-shaped array of ones, or it might consist of all configurations containing as many ones as zeros, or it might be any one of the other 22ab possible subsets of C.)This very large set of possibilities constitutes the class of possible environments e; it is the set of alternatives the adaptive plan must be prepared to handle. The adaptive device's task is to discover or "learn" which element of e is in force by learning what configurations belong to C1.Then, when an arbitrary configuration is presented, the device can reliably indicate whether the configuration belongs to C1, thereby detecting an instance of the pattern. |
|
|
|
|
|