|
Contextualizing Predictive Minds and Machines
Nearly 50 years ago, Allen Newell proposed a largely overlooked
Problem Space Hypothesis, which suggests that the problem space is the fundamental
organizational unit of all human goal-oriented symbolic activity. In recent years, it has been
suggested that humans construct their own contexts or contextual frames to be able to focus on
the task at hand. I will provide some intuition for these propositions. Then, I will introduce
a novel active inference architecture that shows the relevance of a suitable problem space,
that is, a contextual frame, for modeling human behavior. This model derives structural
learning, structural selection, structural parametrization, and planning as inference
processes. The objective is always the same: to minimize anticipated surprise. Considering
challenging reasoning tasks, such as the Abstraction and Reasoning Corpus (ARC) Challenge, I
will finally propose how human problem solving may emerge and may be designed (by evolution)
to reach goals in diverse but locally stable environmental contexts, maximally flexibly and
effortlessly.
|