Some Position Statements for Neuroinformatics-2002

Pei Wang

Department of Computer and Information Sciences, Temple University, USA
webpage: http://www.cis.temple.edu/~pwang/, email: pei.wang@temple.edu

 

Intelligent Control

With AI as the goal, Anokhin's description seems correct, but not concrete and constructive enough to guide research. Furthermore, when the domain is changed from animal to computer, it becomes necessary to distinguish "intelligent control" from "non-intelligent control". It is not clear how the distinction can be made in Anokhin's theory.

Following the tradition of artificial intelligence, we can formalize the control problem similar to the planning/scheduling problem in AI. For a given system with a knowledge representation language L and a set of possible actions A, the system's knowledge on each action can be formalized by a precondition and a postcondition (both are sentences in L), which indicating when it can be invoked and what will be the result of its execution, respectively. In such a framework, the control problem can be formalized as searching for a sequence of actions in A that can turn the initial situation to a final situation, step by step. Then, the action sequence is executed under the system's monitoring. Adjustments are made whenever necessary. Such a description is consistent to Anokhin's theory of functional systems, though in different terminology.

According to my theory on intelligence (see http://www.cogsci.indiana.edu/farg/peiwang/papers.html), what makes intelligent systems different from non-intelligent ones is the capacity of adaptation with insufficient knowledge and resources. Applying this theory to the domain of control, it means an intelligent control system should be able to learn skills (action sequences) from its experience, and to use reasoning and other cognitive functions to estimate the result of an operation in a novel situation, and to adjust plans with respect to the current situation and resources supply.

 

Evolution and Intelligence

The relationship between evolution and intelligence is an important issue, which has not got the attention it deserves.

The relationship can be analyzed on three aspects:

  1. similarity: Both evolution and intelligence are forms of adaptation, which is a process through which a system changes its behaviors according to its environment and certain internal goals. In this process, the system may make various mistakes, but in the long run, the successful adjustments will be kept, and the system's performances will get better and better, with respect it the goals.
  2. differences: Evolution happens in an species, and the changes happens in the "gene" that is passed from generation to generation. The changes are produced independent to the experience of the system, and the good ones are selected through the (survival and reproduction) competition among the individuals of the species. On the other hand, intelligence (according to my theory) happens in an individual, and the change happens in the internal structure of the system, produced according to the experience of the system.
  3. cooperation: To combine the species-level and individual-level adaptation means to have intelligent systems that learn from experience and changes behaviors accordingly, and at the same time, to have the evolution process going on, which produces new (intelligent) systems by coding critical features and parameters as "gene", and selectively keep individual systems according to certain fitness function.

 

The Logic of Reasoning

The Non-Axiomatic Reasoning System (NARS) project (see http://www.cogsci.indiana.edu/farg/peiwang/papers.html) is an attempt to build a logic of inference which provides a unified model for the reasoning, learning, categorization, planning, forecasting, and so on, within an intelligent system.

This theory can be applied to animal, human, and computer, and according to it, the difference in intelligence is basically caused by the different degree of flexibility a system has when changing behaviors according to experience. It can also explain the evolutionary origin of intelligence. For instance, the inference rules used in NARS can be seen as a complicated version of conditioning, which gradually builds new links and nodes (as well as modifies the existing ones) in a network of knowledge.

NARS is being extended to do planning and control.