PlanningEFEMix: Hybrid Active Inference for Sequential Decision-Making under Uncertainty

Abstract

Bhagyeshkumar Chokhawala and Dr. Atif Farid Mohammad

Sequential decision-making under uncertainty remains a central challenge in artificial intelligence and machine learning, especially in environments where agents must act under partial observability, noisy feedback, sparse preference signals, and shifting context. Reinforcement learning, probabilistic planning, and representation learning each provide useful mechanisms for action selection, but each approach also has limitations when deployed as a single decision paradigm. Model-free reinforcement learning can be data intensive and unstable under noise; POMDP-based planning offers principled belief management but depends on accurate transition and observation models; contrastive representation learning improves discrimination but does not by itself define a policy objective; and deterministic Active Inference provides a coherent mechanism for uncertainty reduction and goal satisfaction but is often implemented with a fixed generative model. This paper develops PlanningEFEMix, a hybrid Active Inference planning framework that integrates heterogeneous decision agents using Expected Free Energy (EFE) as a shared meta-level objective. The proposed framework evaluates candidate actions through forward simulation across multiple agents, aggregates agent-specific EFE estimates, and augments action selection with an adaptive state-action bias memory that incorporates experiential feedback from prior outcomes. The algorithm is designed to preserve the interpretability of Active Inference while improving robustness through agent diversity and adaptive policy modulation. A noisy preference inference benchmark is used to frame the experimental evaluation, comparing PlanningEFEMix with deterministic Active Inference, POMDP belief agents, contrastive learning agents, and model-free reinforcement learning baselines across multiple observation- noise regimes and ablation settings. The manuscript contributes a formal problem definition, an algorithmic specification, an architectural framework, an evaluation protocol, and a discussion of deployment implications for decision-making systems that require explainability, robustness, and adaptive behavior under uncertainty.

PDF