Overview
A new framework for combinatorial optimization has been introduced, designed to operate on natively analog, unconventional computing platforms. This framework utilizes sum-constrained continuous quadratic programs, which are solvable by QCi's Dirac-3S photonic entropy computer. The foundation of this approach lies in the Motzkin-Straus theorem, which establishes a connection between discrete clique problems and optimization challenges over the probability simplex. The research focused on demonstrating the framework's adaptability by addressing constraint satisfaction problems.
Research Context
Combinatorial optimization problems often involve navigating complex, non-convex landscapes to identify optimal or near-optimal solutions. Traditional computing methods face challenges in efficiently solving certain classes of these problems. The development of unconventional computing platforms, such as photonic entropy computers, represents an alternative approach. This work specifically positions entropy computing as a competitive methodology for tackling non-convex optimization problems, while also aiming to provide rigorous baselines for this emerging computational paradigm.
Approach
The core of the approach involves translating discrete clique problems into sum-constrained continuous quadratic programs. This translation is enabled by the Motzkin-Straus theorem, which provides the mathematical bridge for this transformation. The continuous quadratic programs are then executed and solved using QCi's Dirac-3S photonic entropy computer. To assess its performance and versatility, the framework was applied to solve constraint satisfaction problems. Extensive benchmarks were conducted utilizing the DIMACS suite, a standard collection of benchmark instances for various combinatorial optimization problems.
Findings
The Dirac-3S photonic entropy computer, operating within this new framework, exhibited significant performance characteristics when benchmarked against two independently implemented classical baselines on the DIMACS suite:
- Overall Performance: The Dirac-3S platform matched or surpassed the performance of both classical baselines on more than four-fifths (over 80%) of the benchmark instances tested.
- Structured Graph Families: On nearly all structured graph families within the DIMACS suite, the photonic entropy computer achieved the best known solution.
- Large Instances: For several of the largest instances tested, the Dirac-3S platform notably outperformed both classical solvers.
- Challenging Instances: Well-tuned classical continuous optimizers maintained an advantage specifically on the hardest planted-clique instances.
Why This Matters
This work establishes a viable pathway for solving combinatorial optimization problems using natively analog unconventional computing platforms. By demonstrating competitive performance against classical methods, particularly on structured graph families and larger instances, the research positions entropy computing as a noteworthy approach for navigating non-convex landscapes. It also contributes to the establishment of rigorous baselines for an emerging computational paradigm.