Intrinsic motivation as constrained entropy maximization

TL;DR

This paper presents a unified perspective on intrinsic motivation in intelligent systems by framing it as constrained entropy maximization, linking active inference, empowerment, and free energy concepts.

Contribution

It introduces a general framework connecting various intrinsic motivation theories through constrained maximum entropy inference, offering new insights into their relationships.

Findings

Active inference, empowerment, and intrinsic motivation are unified under constrained entropy maximization.

The connection between free energy and empowerment is further elucidated.

Implicit model-evidence constraints are identified in maximum-occupancy approaches.

Abstract

"Intrinsic motivation" refers to the capacity for intelligent systems to be motivated endogenously, i.e. by features of agential architecture itself rather than by learned associations between action and reward. This paper views active inference, empowerment, and other formal accounts of intrinsic motivation as variations on the theme of constrained maximum entropy inference, providing a general perspective on intrinsic motivation complementary to existing frameworks. The connection between free energy and empowerment noted in previous literature is further explored, and it is argued that the maximum-occupancy approach in practice incorporates an implicit model-evidence constraint.