|
|
|
|
|
|
|
The unit(s) implementing the detector, when activated, would broadcast a signal with an identifying prefix. (For the reader familiar with the early history of pattern recognition these units act much like the demons at the lowest level of Selfridge's Pandemonium [1959].) Other devices would "weight" the signals, "sum" them, and "compare" the result to a threshold (cf. section 7.3) to determine which response signal (from the set of transformations {hi}) should be broadcast. More generally, the behavioral atoms would be a string of broadcast units with an ''initiate" condition C which specifies the set of signals capable of activating the atom, an "end" signal J which indicates the end of the atom's activity, and a "predicted value" signal which is meant to indicate the ultimate value to the behavioral unit of that atom's activation. With this arrangement we can treat the behavioral unit as a population of atoms. The atom activated at any given time is determined by a competition between whatever atom is already activated and all atoms having condition C satisfied by a signal in the current state S(t). The higher the predicted value of the atom the more likely it is to win the competition. The object at this level is to have each atom's predicted value V consistent with that of its successor, so that a set of atoms acting in sequence provides a consistent prediction of their value to the behavioral unit. (In this way the atoms satisfy the error correction requirements discussed at the beginning of section 3.4 under element (iii) of a typical search plan.) This object can be accomplished via a genetic plan applied to the population of atomsthe reproduction of any atom is determined by the match between its predicted value and that of whatever atom is next activated. For example, consider two atoms, a1 with parameters (C, J, V) and a2 with parameters (C', J', V'), where a1's end signal acts as a2's initiate signal. Then (V' - |V - V'|) could be used as a payoff to a1 for purposes of the genetic plan, since the quantity measures the match between V' and V. The population would then be modified as outlined in section 6.1, new atoms being assigned the predicted value of the successor a2. That is, the offspring of a1 would be assigned the predicted value V'. All atoms active since the last actual payoff from the environment, and their offspring, are tagged and their predicted values are adjusted up or down at the time of the next environmental payoff. The adjustment is determined by the difference between predicted value and the actual payoff rate (payoff received from the environment divided by the elapsed time since last payoff). After each environmental payoff the active behavioral unit is subjected to a genetic plan (again as described in section 6.1). The behavioral unit next active (after the environmental payoff) is determined by the winner of a competition among all atoms in all behavioral units. The outcome of the competition is determined in the same way as the within-unit competition. Finally, a behavioral unit may be subjected to |
|
|
|
|
|