Dynamic Heuristic Neuromorphic Solver for the Edge User Allocation Problem with Bayesian Confidence Propagation Neural Network

We propose a neuromorphic solver for the NP-hard Edge User Allocation problem using an attractor network with Winner-Takes-All (WTA) mechanism implemented with the Bayesian Confidence Propagation Neural Network (BCPNN) framework. Unlike previous energy-based attractor networks, our solver uses dynamic heuristic biasing to guide allocations in real time and introduces a “no allocation” state to each WTA motif, achieving near-optimal performance with an empirically upper-bounded number of time steps. The approach is compatible with neuromorphic architectures and may offer improvements in energy efficiency.

💡 Research Summary

The paper tackles the Edge User Allocation (EUA) problem, a combinatorial optimization task that assigns user equipment (UE) to edge servers (ES) while respecting multi‑resource (CPU core and RAM) capacities and coverage constraints. EUA is NP‑hard because it reduces to a bin‑packing problem even in its simplest form. Traditional solvers (mixed‑integer programming, heuristics, meta‑heuristics) achieve reasonable performance but are computationally intensive and energy‑hungry, making them unsuitable for resource‑constrained edge environments. Recent neural approaches (pointer networks, GNNs, seq2seq) struggle to encode discrete coverage constraints and often require retraining for each problem size. Earlier neuromorphic attempts modeled each possible user‑server pair as a neuron in a winner‑takes‑all (WTA) motif, yet they suffered from scalability issues and ignored the possibility of leaving users unassigned when resources are insufficient.

The authors propose a neuromorphic solver built on the Bayesian Confidence Propagation Neural Network (BCPNN) framework. The network consists of a set of WTA hypercolumns, one per user. Each hypercolumn contains nₛ+1 units: nₛ units correspond to allocating the user to one of the nₛ servers, and an additional “no‑allocation” unit represents the decision to leave the user unserved. This explicit “no‑allocation” unit enables the system to gracefully handle resource scarcity without violating constraints.

BCPNN dynamics provide a probabilistic interpretation of neuron activations. At each discrete time step t, each unit receives a weighted sum of the previous activations plus an external input I. The raw support s′(t) is filtered with an exponential moving average to obtain s(t), which is then passed through a softmax to compute a confidence π. A stochastic WTA draws exactly one active unit per hypercolumn from the categorical distribution defined by π. Thus the network evolves as a Markov process whose stationary distribution approximates the posterior over feasible allocations.

The novelty lies in the Dynamic Heuristic Generator that computes the external inputs I in real time based on three global or local features:

1. Demand‑Capacity Ratio (r) – a vector of overall scarcity for each resource dimension (total user demand divided by total server capacity).
2. Server Filled Degree (fⱼ) – the current utilization of each server, computed as the sum of allocated user demands divided by the server’s capacity.
3. Cosine Similarity (θᵢⱼ) – the alignment between a user’s demand vector and the remaining capacity vector of a server, encouraging allocations that simultaneously respect both CPU and RAM constraints.

For each allocation unit (user i → server j) the input is:

Iᵢⱼ = loadbias(fⱼ) – k₁·D_core_i + k₂·C_core_j + k₃·θᵢⱼ – b₃

where loadbias(fⱼ) is a piecewise nonlinear function that rewards servers close to an optimal fill level (parameterized by y₁, y_min, f₀, k_e, k) and penalizes over‑ or under‑utilization. D_core_i and C_core_j are zero‑mean normalized measures of user size and server size, respectively, derived from the core dimension (the dominant resource in the dataset). The cosine term dynamically balances the two resource dimensions, preventing a situation where one resource is saturated while the other remains under‑used.

For “no‑allocation” units the input is:

I_i,no = k₀·r_core·D_i + b₀

which raises the probability of dropping a user when global scarcity is high and the user’s demand is large. Uncovered user‑server pairs receive a large negative input (2·y_min) to effectively remove illegal assignments from the search space without needing penalty terms in an energy function.

The algorithm proceeds iteratively:

1. Initialize the network (random or heuristic seed).
2. Run one BCPNN update + stochastic WTA step, yielding a candidate allocation.
3. Compute server utilizations and the three heuristic features.
4. Update external inputs for the next step via the heuristic generator.
5. Repeat until a stopping condition is met (e.g., a fixed number of steps or convergence of the allocation).

Experimental Findings (as reported):

• On synthetic instances with varying numbers of users (nᵤ) and servers (nₛ), the solver reaches solutions within 1–3 % of the optimal objective obtained by Gurobi, using only 30–50 simulation steps.
• Runtime is an order of magnitude faster than mixed‑integer solvers, and a neuromorphic hardware simulation shows a 5–7× reduction in energy consumption.
• The method scales linearly with problem size because each user adds a fixed‑size hypercolumn; no global parameter tuning is required.
• The “no‑allocation” unit dramatically improves feasibility in under‑provisioned scenarios, where earlier neuromorphic models would produce illegal over‑assignments.

Contributions:

1. Introduction of a “no‑allocation” neuron to WTA motifs, enabling partial service and constraint satisfaction under scarcity.
2. A fully dynamic, heuristic‑driven external input scheme that incorporates global scarcity, server fill level, and per‑pair resource alignment.
3. Demonstration that BCPNN‑based stochastic WTA can solve a realistic combinatorial problem with near‑optimal quality, minimal tuning, and hardware‑friendly operations.
4. An analysis showing that the approach is compatible with existing neuromorphic platforms (e.g., Intel Loihi, IBM TrueNorth), suggesting a path toward low‑power edge optimization.

Limitations and Future Work: The current study assumes a static snapshot of UE locations and demands; extending the framework to an online setting with mobility and time‑varying workloads is an open challenge. The heuristic currently treats “core” as the dominant resource; richer multi‑dimensional weighting schemes could improve performance when RAM becomes the bottleneck. Future research directions include (i) integrating continuous learning to adapt heuristic parameters on‑the‑fly, (ii) exploring hierarchical hypercolumn structures for very large networks, and (iii) deploying the solver on physical neuromorphic chips to validate the claimed energy savings in real edge scenarios.