Imported from ActiveInferenceInstitute/courses (course_development/active_inference_101/04_implementation/08_planning/AGENTS.md). Install upstream with npx skills add ActiveInferenceInstitute/courses --skill 08_planning. Copyright stays with the author.

Station: Planning (Implementation & Simulation)

Quick Navigation: Module README | Course AGENTS

Conventions

• Perspective: Python, pymdp, agent-based modeling
• Topics: Complete Active Inference agent, multi-step planning, sophisticated inference implementation, performance benchmarking, extension projects
• Lab Style: Coding Assignment
• Audience: College 1st semester undergraduates
• Tone: Rigorous but accessible

Content Guidelines

This module brings together all components into a complete Active Inference agent with planning capabilities. Content should:

1. Implement multi-step planning: Extend EFE evaluation over multiple future time steps, evaluating policy trees rather than single actions.
2. Define key implementation concepts:
   • Policy tree enumeration: Generating all possible action sequences up to a given planning horizon and evaluating each via EFE
   • Sophisticated inference: A recursive implementation where the agent simulates its own future belief updating when evaluating policies
   • Computational budget: Strategies for limiting the number of policies evaluated (pruning, beam search, habit-based shortcuts)
   • Performance benchmarking: Comparing agent performance across planning horizons, precision settings, and with/without learning
3. Run the complete agent: Execute a full simulation with perception, action selection, learning, and multi-step planning in a complex environment.
4. Suggest extension projects: Provide ideas for students to extend the agent (new environments, continuous state spaces, hierarchical models).

Active Inference Integration

• This module is the capstone: all Active Inference components (generative model, state inference, EFE, parameter learning) are integrated
• Sophisticated inference implements the most advanced form of planning from the Mathematical Frameworks course
• Performance benchmarking connects computational implementation back to theoretical predictions

Assessment Alignment

Questions should test the ability to:

• Implement multi-step EFE evaluation for a planning horizon of 2-3 steps
• Compare agent performance with different planning depths
• Identify computational bottlenecks and propose solutions

Ensure all content adheres to ../../resources/notation_table.md.