Active Inference Agent (Generative Model + EFE + Empowerment)

What problem does it solve?

The Active Inference Agent enables autonomous decision-making in uncertain environments by unifying a generative-world-model with a principled planning objective (Expected Free Energy) to minimize surprise while balancing goal achievement, information gathering, and controllability.

Core Features & Use Cases

• Modular four-component world model: Latent Encoder q(s|o), Likelihood P(o|s), Transition P(s'|s,a), and Preferences C.
• Three-term EFE planning: Pragmatic, Epistemic, and Instrumental terms with horizon normalization, per-step terms, and FP32 precision invariants.
• Flexible planning engines: Random Shooting, Cross-Entropy Method (CEM), and an Amortized Policy for fast inference, with optional discrete POMDP backends via pymdp.
• Offline RL integration: Minari-based dataset training and evaluation, plus telemetry for model debugging and optimization.
• Research-to-deployment bridge: end-to-end workflow from data to planning, including configuration presets for small and large-scale deployments.

Quick Start

Create a tiny Active Inference agent with a minimal config and call reset, then plan from a small observation to obtain an action.