A perspective on artificial intelligence: Learning to learn

The classical approach to the acquisition of knowledge in artificial intelligence has been to program the intelligence into the machine in the form of specific rules for the application of the knowledge: expert systems. Unfortunately, the amount of time and resources required to program an expert system with sufficient knowledge for non-trivial problem-solving is prohibitively large. An alternative approach is to allow the machine tolearn the rules based upon trial-and-error interaction with the environment, much as humans do. This will require endowing the machine with a sophisticated set of sensors for the perception of the external world, the ability to generate trial actions based upon this perceived information, and a dynamic evaluation policy to allow it to measure the effectiveness of its trial actions and modify its repertoire accordingly. The principles underlying this paradigm, known ascollective learning systems theory, have already been applied to sophisticated gaming problems, demonstrating robust learning and dynamic adaptivity.The fundamental building block of a collective learning system is thelearning cell, which may be embedded in a massively parallel, hierarchical data communications network. Such a network comprising 100 million learning cells will approach the intelligence capacity of the human cortex. In the not-too-distant future, it may be possible to build a race of robotic slaves to perform a wide variety of tasks in our culture. This goal, while irresistibly attractive, is most certainly fraught with severe social, political, moral, and economic difficulties.This paper was given as an invited talk on the 12th Symposium on Operations Research, University of Passau, September 1987.