Principia Cybernetica Web

Learning and Model-Building

Cybernetic epistemology is in essence constructivist: knowledge cannot be passively absorbed from the environment, it must be actively constructed by the system itself. The environment does not instruct, or "in-form", the system, it merely weeds out models that are inadequate, by killing or punishing the system that uses them. At the most basic level, model-building takes place by variation-and-selection or trial-and-error. Let us illustrate this by considering a primitive aquatic organism whose control structure is a slightly more sophisticated version of the thermostat. To survive, this organism must remain in the right temperature zone, by moving up to warmer water layers or down to colder ones when needed. Its perception is a single temperature variable with 3 states X = {too hot, too cold, just right}. Its variety of action consists of the 3 states Y = {go up, go down, do nothing}. The organism's control knowledge consists of a set of perception-action pairs, or a function f: X ;->Y. There are 33 = 27 possible such functions, but the only optimal one consists of the rules:
  • too hot ;-> go down
  • too cold -> go up
  • just right -> do nothing.
The last rule could possibly be replaced by either just right -> go up or just right -> go down. This would result in a little more expenditure of energy, but in combination with the previous rules would still keep the organism in a negative feedback loop around the ideal temperature. All 24 other possible combinations of rules would disrupt this stabilizing feedback, resulting in a runaway behavior that will eventually kill the organism.

Imagine that different possible rules are coded in the organism's genes, and that these genes evolve through random mutations each time the organism produces offspring. Every mutation that generates one of the 24 combinations with positive feedback will be eliminated by natural selection. The three negative feedback combinations will initially all remain, but because of competition, the most energy efficient combination will eventually take over. Thus internal variation of the control rules, together with natural selection by the environment eventually results in a workable model.

Note that the environment did not instruct the organism how to build the model: the organism had to find out for itself. This may still appear simple in our model with 27 possible architectures, but it suffices to observe that for more complex organisms there are typically millions of possible perceptions and thousands of possible actions to conclude that the space of possible models or control architectures is absolutely astronomical. The information received from the environment, specifying that a particular action or prediction is either successful or not, is far too limited to select the right model out of all these potential models. Therefore, the burden of developing an adequate model is largely on the system itself, which will need to rely on various internal heuristics, combinations of pre-existing components, and subjective selection criteria to efficiently construct models that are likely to work.

Natural selection of organisms is obviously a quite wasteful method to develop knowledge, although it is responsible for most knowledge that living systems have evolved in their genes. Higher organisms have developed a more efficient way to construct models: learning. In learning, different rules compete with each other within the same organism's control structure. Depending on their success in predicting or controlling disturbances, rules are differentially rewarded or reinforced. The ones that receive most reinforcement eventually come to dominate the less successful ones. This can be seen as an application of control at the metalevel, or a metasystem transition, where now the goal is to minimize the perceived difference between prediction and observation, and the actions consist in varying the components of the model.

Different formalisms have been proposed to model this learning process, beginning with Ashby's homeostat, which for a given disturbance searched not a space of possible actions, but a space of possible sets of disturbance -> action rules. More recent methods include neural networks and genetic algorithms. In genetic algorithms, rules vary randomly and discontinuously, through operators such as mutation and recombination. In neural networks, rules are represented by continuously varying connections between nodes corresponding to sensors, effectors and intermediate cognitive structures. Although such models of learning and adaptation originated in cybernetics, they have now grown into independent specialisms, using labels such as "machine learning" and "knowledge discovery".

Heylighen F. & Joslyn C. (2001): "Cybernetics and Second Order Cybernetics", in: R.A. Meyers (ed.), Encyclopedia of Physical Science & Technology , Vol. 4 (3rd ed.), (Academic Press, New York), p. 155-170

Copyright© 2001 Principia Cybernetica - Referencing this page

F. Heylighen, & C. Joslyn,

Sep 3, 2001


Metasystem Transition Theory


Prev. Next


Add comment...