Fully Parallel Multi-Agent Photonic Optimizer

TL;DR

The paper proposes an integrated photonic optimizer (IPO) that performs in-memory, parallel matrix–vector computations across multiple wavelength channels to solve optimization problems with a cooperative multi-agent framework. It leverages wavelength-division multiplexing and photonic crossbars to encode each agent as a distinct optical channel, enabling truly parallel MAC operations directly where data reside. The authors demonstrate proof-of-concept hardware and algorithms on discrete Max-Cut tasks and continuous linear/programming-style problems, highlighting robustness to hardware noise and potential for high compute density and energy efficiency. By treating hardware noise as a computational resource and scaling the number of agents, the approach enables rapid exploration and convergence toward high-quality solutions in both combinatorial and continuous domains.

Abstract

Optimization problems are central to many important cross-disciplinary applications.In their conventional implementations, the sequential nature of operations imposes strict limitations on the computational efficiency. Here, we discuss how analog optical computing can overcome this fundamental bottleneck. We propose a photonic optimizer unit, together with supporting algorithms that uses in memory computation within a nature inspired, multi agent cooperative framework. The system performs a sequence of reconfigurable parallel matrix vector operations, enabled by the high bandwidth and multiplexing capabilities inherent to photonic circuits. This approach provides a pathway toward fast paced and high quality solutions for difficult optimization and search problems.

Figures (6)

• Figure 1: Concept of Photonic Optimizer. (a) Illustration of an multi-agent optimization. A foraging avian must explore ($\vec{d}$) the landscape using their characteristic skills (i.e. sensory range). The likelihood of finding the optimal food resource is improved through cooperation with a population of avians. When translated to the energy landscape of a typical NP-hard optimization problem, this strategy, leveraging the individual and social characteristics of multiple independent agents, can enable convergence to the global optimum in difficult optimization tasks. (b) Comparison of digital electronic architectures with the proposed photonic optimizer for solving multi-agent optimization problems. Digital electronics require many sequential processing steps, which can be executed on a single core or distributed across multiple cores. In contrast, the entire problem can be processed within a single photonic optimizer in a truly parallel, in-memory computing manner. (c) Flowchart describing the operational flows in the proposed integrated optimizer unit. When properly defined through chosen constraints, loss functions and update rules, the optimizer can iterate over a problem with multiple agents, evaluating all agents in a single cycle.
• Figure 2: Conceptual Integrated Photonic Optimizer Unit. (a) In an IPO unit, every agent is encoded as an input vector carried on a unique wavelength (optical carrier). The optimization problem is then represented and solved through matrix–vector operations, with the matrix elements encoded in the states of photonic memory devices. Multiple agents can be processed in parallel, leveraging both in-memory and recurrent computations to efficiently explore the solution space. (b) Conceptual schematic of the compute core. The core consists of multiple individually addressable wavelength channels (one per agent) and a high optical bandwidth linear processor. After demultiplexing a broadband light source (e.g., ASE source, laser array, or frequency comb), each wavelength channel is split into $m$ distinct paths, each modulated by a tunable attenuator (TA) (e.g., EAM). This allows an input vector of length $m$ to be encoded on each channel. The channels are then multiplexed into the $m$ input waveguides of a photonic crossbar array, with each intersection also loaded with a tunable attenuator (e.g., phase-change material, EAM). The crossbar performs multiply-and-accumulate operations between matrix elements $w_{ij}$ and all frequency channels in a single pass. The resulting vectors of are measured via demultiplexed photodetectors. The outputs of the photonic agent evaluation are then processed according to the chosen algorithm to update the inputs for the next cycle.
• Figure 3: Prototype Integrated Photonic Optimizer Unit. a The conceptual IPO is evaluated for functionality and performance through a prototype hardware. Four distinct C-band wavelength channels (200GHz ITU grid channels C28, C30, C32 and C34) are spatially multiplexed to perform matrix-vector operations. A high-speed FPGA is interfaced with the photonic chip to handle auxiliary operations. Light from a broadband ASE source is split by a demultiplexer into the select channels, each subsequently amplified in two stages to roughly [100]mW before entering the silicon PIC. The PIC incorporates programmable input weights along with a 9×3 programmable crossbar array. Photocurrents from the on-chip photodetectors are amplified by off-chip TIAs and fed back into the same FPGA, which also controls the input signals and matrix weighting. b A plot depicting the spectrum of the ASE source along with the four distinct 200 GHz channels. c A plot showing L2-error of matrix-vector multiply operations for each input channel across averaging, with fitted noise profiles obtained through average error of 1280 random input vectors length 64 and a constant matrix weights.
• Figure 4: Combinatorial Optimization Problems. (a) The top panel is an illustration showing the IPO approach to sample a problem's solution landscape and provide convergence to the optimal solution through iterative changes. The inset shows the decrease in the system’s Shanon entropy as the network approach convergence. The bottom panel shows the configuration of the photonic crossbar. The problems can be solved deterministically, or through stochastic and evolutionary processes. (b) Exemplary emulated optimization of a circular graph of 40 nodes using measured compute accuracy levels for the four ITU-channels from the benchmark measurement. The inset is a cartoon of the cyclic graph optimization problem Max-cut. The dashed red line cuts the graph into two sets of complementary nodes. The edges are encoded as matrix weights in the photonic crossbar. (c) Top Panel: Average global energy of twenty runs discovering the Max-cut of a circular graph of size 40 for different decaying noise profiles. The cyan line shows the average for purely deterministic calculations. The red line the performance using experimentally measured noise levels. While deterministic multi-agent approach can enable optimal convergence with increasing number of agents and iterations, we find optimal noise profiles to allow quicker convergence to the optimal solutions. (d) Performance Benchmarks on non circular graphs. Examples are taken from the Biq-Mac library biqmac. We use the experimental noise profiles and four agents to find solutions for each graph 30 times with different initial conditions. Even without deliberate hyperparameter tuning, all graphs are solvable. The bottom panel shows the success rate, while the top panel displays the distribution of relative errors in percent. In the top panel, the bars indicate the maximum, minimum, and mean relative error.
• Figure 5: Linear Programming Problems. (a) A schematic showing the configuration of a crossbar unit used for solving a constrained linear problem through canonical particle swarm approach. Multiple queries encoded into the amplitude modulated optical signals operate in parallel. Evaluation of constraints are uniquely handled in the same pass by mapping into extra columns in the crossbar. (b) A snapshot of 6 agents’ trajectories during balanced exploration and exploitation under moderate injected photonic noise in an emulated particle swarm optimization problem. The bottom panel shows evolution of the cost function as a function of iterations. (c) A plot showing the final average swarm position distance from the true minimum for canonical particle swarm problem for swarm sizes 2 to 16. The experimentally achievable noise levels and comparative approximate bit-accuracies are indicated.(d) The minimum swarm loss as a function of number of iterations for a swarm size 10 is shown. Here, the perceived loss is the loss recorded by the swarm through noisy evaluations, while true loss the exact loss at the resulting positions of each agent.