BRAIN: Bayesian Reasoning via Active Inference for Agentic and Embodied Intelligence in Mobile Networks

TL;DR

BRAIN harnesses a deep generative model of the network environment and minimizes variational free energy to unify perception and action in a single closed-loop paradigm and exhibits real-time interpretability of its decisions through human-interpretable belief state diagnostics.

Abstract

Future sixth-generation (6G) mobile networks will demand artificial intelligence (AI) agents that are not only autonomous and efficient, but also capable of real-time adaptation in dynamic environments and transparent in their decisionmaking. However, prevailing agentic AI approaches in networking, exhibit significant shortcomings in this regard. Conventional deep reinforcement learning (DRL)-based agents lack explainability and often suffer from brittle adaptation, including catastrophic forgetting of past knowledge under non-stationary conditions. In this paper, we propose an alternative solution for these challenges: Bayesian reasoning via Active Inference (BRAIN) agent. BRAIN harnesses a deep generative model of the network environment and minimizes variational free energy to unify perception and action in a single closed-loop paradigm. We implement BRAIN as O-RAN eXtended application (xApp) on GPU-accelerated testbed and demonstrate its advantages over standard DRL baselines. In our experiments, BRAIN exhibits (i) robust causal reasoning for dynamic radio resource allocation, maintaining slice-specific quality of service (QoS) targets (throughput, latency, reliability) under varying traffic loads, (ii) superior adaptability with up to 28.3% higher robustness to sudden traffic shifts versus benchmarks (achieved without any retraining), and (iii) real-time interpretability of its decisions through human-interpretable belief state diagnostics.

Figures (8)

• Figure 1: Preliminary experiment on catastrophic forgetting issue for agent in network slicing. (a) Sequential distribution shifts: agent trains initially on Slice 1 (eMBB, green), then shifts to Slice 2 (URLLC, red), and finally Slice 3 (mMTC, blue), each with distinct QoS requirements and traffic patterns. (b) The impact of distribution shifts on performance. Initially, at ① the agent learns Slice 1 effectively. When data distribution transitions at ② to Slice 2, the agent learns new policies while starting to forget previously acquired knowledge (Slice 1). At ③, another shift to Slice 3 occurs, further reducing performance on previously learned Slices 1 and 2. Eventually, at ④, the scenario cycles back, requiring costly retraining due to performance degradation caused by catastrophic forgetting.
• Figure 2: Architecture of conventional DRL and proposed explainable deep active inference agents.
• Figure 3: Overview of GPU-Accelerated Testbed.
• Figure 4: Training dynamics of agentic (RL/DRL) and embodied (active-inference) Agents on AI-RAN testbed: i) Mean cumulative reward per episode (higher is better), showing convergence speed and asymptotic control performance. ii) Average training loss (lower is better), used as a stability proxy for optimization dynamics during online learning. iii) Policy entropy (higher indicates more exploration), capturing the exploration-exploitation evolution over training.
• Figure 5: Agent’s posterior belief trajectory over hidden traffic demand levels ( Low, Medium, High) for each network slice across episodes. Time on the x-axis indexes discrete decision epochs (e.g., observation-update intervals $\times 10^3$), and the y-axis enumerates the three demand states ( Low at bottom to High at top). Color encodes the belief probability $Q(s_t)$ (brighter/yellow = higher, darker/purple = lower). White arrows mark the agent’s explicit information-gathering Check actions.