Exploring Active Inference for Intelligent Resource Allocation in Emergent Wireless Networks

Future wireless networks are anticipated to support a wide range of intelligent and human centric applications, including the Internet of Things (IoT), smart cities, autonomous driving, and the metaverse. Consequently, there is an increasing need to enhance their cognitive and decision-making capabilities to autonomously optimize network resources. Specifically, unmanned aerial vehicles (UAVs) and cognitive non-orthogonal multiple access (Cognitive NOMA) are being integrated for their potential to make real-time decisions while tackling limited spectrum resources and supporting massive connectivity. Maximizing the sum rate of a UAV-based Cognitive-NOMA system through joint optimization of subchannel assignment, power allocation, and UAV mobility presents a highly nonconvex and NP-hard problem due to the joint optimization variables and the explosive search space. This challenge is often formulated as an optimization problem using random variables. However, traditional approaches to finding optimal solutions typically require iterative or exhaustive searches over all possible combinations of subchannel assignment, power allocation, and UAV mobility, leading to excessive computational complexity. Furthermore, machine learning models for wireless resource optimization, often trained on datasets that fail to capture the full complexity of real-world scenarios, struggle to handle nonstationary events effectively. To address this nonconvex optimization problem, this thesis explores a novel data-driven, active inference-based learning framework inspired by cognitive neuroscience. Compared to classical numerical optimization algorithms and machine learning models, which are not designed for causal reasoning and interpretability, the proposed active inference framework enables agents to interact with their environment proactively, dynamically extract new data, and make real-time decisions. The problem is formulated as a prediction error minimization using a unique Active Generalized Dynamic Bayesian Network (Active-GDBN) that operates in both discrete and continuous states, facilitated by constant neuronal message passing. The main idea is to process the unknown nonlinear input from an exhaustive search optimization algorithm (expert domain knowledge) using the Active-GDBN framework. In the offline training phase, the UAV utilizes solutions derived from the exhaustive search strategy to acquire knowledge and build a dynamic generative model that captures the complex relationships and dependencies among subchannel assignments, power distributions, and UAV mobility. During the online active inference phase, the UAV leverages the trained generative model. It dynamically selects actions such as discrete subchannels and continuous power allocation based on its position to maximize the sum rate. To further enhance the cognitive capabilities of the proposed solution, we integrate emergent semantic context-aware reasoning, enabling the design of a more intelligent and adaptable resource allocation scheme. Numerical simulations validate the effectiveness of the proposed solution, revealing comparable performance near the optimal exhaustive search and demonstrating superiority over other feasible standard benchmark approaches in terms of achievable sum rates.