A Neuromorphic Solver for the Edge User Allocation Problem with Bayesian Confidence Propagation Neural Network : A Dynamic Heuristic Generator for External Unit Excitation

Sammanfattning

Edge computing pushes computation from remote clouds to resourceconstrained servers close to end-users. Determining which users should be connected to which edge servers—the Edge User Allocation (EUA) problem —is NP-hard and becomes intractable for large instances when solved by conventional mixed-integer programming. While many approximate methods have been proposed, neuromorphic solutions are particularly appealing due to their potential for high-speed and energy-efficient hardware implementation. These approaches leverage parallel, stochastic dynamics and local competition to explore combinatorial spaces and naturally enforce exclusivity constraints. In this thesis, we present a neuromorphic formulation based on the Bayesian Confidence Propagation Neural Network (BCPNN), in which each user is represented by a winner-take-all module composed of neurons corresponding to possible user-server assignments, including a dedicated unit for the “no allocation” option. Instead of encoding all constraints in a static energy function, we introduce a dynamic bias generator that steers the network with three heuristics: (i) a load-bias curve that favours near-full servers while penalizing overutilization and underutilization, (ii) a size heuristic that prioritizes smaller-demand users and larger-capacity servers, and (iii) a cosine-similarity term that matches a user’s demand vector to the current residual capacity of each server, enforcing dual-resource feasibility online. A single global parameter controls the trade-off between the number of active servers and the number of served users; scanning a small grid of values enables exploration of different levels of user-server tradeoff. Experiments on a 30-instance synthetically generated benchmark, each accompanied by an optimal Gurobi solution, show that the proposed BCPNNEUA solver finds feasible allocations whose score lies within an average of 12.6% of the optimum while converging in a few hundred simulation steps. Because the algorithm relies on local updates and event-driven communication, it is well suited to low-power neuromorphic hardware, offering a scalable and energy-efficient alternative for real-time edge-resource management.