Divide et impera: subgoaling reduces the complexity of probabilistic inference and problem solving

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Humans and animals naturally use subgoaling to manage complex problems.
• Mechanisms for online subgoal selection in novel problem-solving remain unclear.
• Optimal conditions and strategies for subgoaling require further investigation.

Purpose of the Study:

• To present a computational framework for understanding subgoal selection in problem-solving.
• To propose that effective subgoals minimize information resource use for parsimonious inference and control.
• To explain the adaptive advantages of subgoaling for cognitive processes.

Main Methods:

• Developed a computational model based on Occam's razor principle for subgoal selection.
• Implemented the model using approximate probabilistic inference and a sampling method considering descriptive complexity.
• Validated the model using a reinforcement learning benchmark (four-rooms scenario).

Main Results:

• The proposed method demonstrated reduced inferential steps compared to non-subgoaling approaches.
• The model selected more compact control programs, indicating enhanced efficiency.
• The computational framework provided a mechanistic explanation for prefrontal cortex neuronal dynamics in planning tasks.

Conclusions:

• Subgoaling, guided by information parsimony, enhances planning, control, and learning.
• This approach offers a unified perspective on the adaptive benefits of subgoaling.
• The model reduces cognitive effort and working memory load during complex problem-solving.