Goals and the Structure of Experience

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Neuroscience

Background:

• Purposeful behavior in natural and artificial intelligence is often linked to world models.
• Traditional models separate descriptive (state representation) and prescriptive (reward function) aspects.
• An alternative, computationally unexplored, is the interdependent co-emergence of these aspects from goals.

Purpose of the Study:

• To propose a novel computational framework for goal-directed state representation in cognitive agents.
• To demonstrate how descriptive and prescriptive world model aspects can co-emerge from agent-environment interactions.
• To introduce the concept of 'telic states' for a unified understanding of purposeful behavior.

Main Methods:

• Development of a computational framework based on agent-environment interaction sequences (experiences).
• Introduction of 'telic states' defined as classes of goal-equivalent experience distributions.
• Leveraging Buddhist epistemology to define goal-directed states.

Main Results:

• The framework shows how descriptive and prescriptive world model components can interdependently co-emerge from an agent's goal.
• Telic states offer a parsimonious account of goal-directed learning.
• This approach reinterprets goal-directed learning via statistical divergence between policies and desired experiences.

Conclusions:

• The proposed framework provides a novel computational perspective on purposeful behavior.
• Telic states offer a unified account integrating behavioral, phenomenological, and neural dimensions.
• This approach has potential applications across diverse intelligent systems, both natural and artificial.