Goal-Oriented Intelligence: A Workable Hypothesis?

In cognitive science, one starts from the assumption that cognitive functions are, or at least can be modeled by computations. Then, we need a pragmatic definition for intelligence that lends itself into a workable algorithm. We start from the hypothesis that basically all facets of intelligence are related to goal oriented behavior. Goal oriented behavior, however, can be the result of evolution and may not be intelligent per se. On the other hand, intelligence can manifest itself through communication. We consider problem types of different complexities and (i) establish the category of problems that are worth to communicate, (ii) give a definition for intelligence based on this special category, and (iii) identify another computational problem type, which is necessary for communication and which is highly problematic for present day machine learning algorithms.

Communication requires agreements about symbol meaning associations. We show that such agreements are very hard without a mind model, where mind simply means a predictive model of the communicating partner and partial access (observation) to her actual internal rewards (emotions). 

We will present two examples to illustrate matters. Our project called “Innovation Engine in BlogSpace” intends to develop information seeking conversational agents that could interfere with people in BlogSpace. The other example is about “Testing and communicating with severely handicapped, non-speaking, but speech understanding children”, where the goal is to estimate the zone of proximal development and to optimize training materials. Very recent results on collaborative filtering made recommendation systems highly efficient provided that databases are available. Collection of the data without endangering privacy has become feasible.