Cognitive Raman Amplifier Control Using an Evolutionary Optimization Strategy

Online versus Offline Optimization Methods for Raman Amplifier Optimization

“…once they are ready, ultra-fast RA gain optimization can be achieved, relying on a simple and fast matrix if just NN inv is applied, or taking just a few milliseconds with GD + NN fwd . In contrast, the proposed DE can be reusable for any RA experimental setup with minimum time and effort, not requiring physical layer characterization as in an offline approach [14]. Nevertheless, DE's flexibility and highly accurate solutions come with the cost of having a very long optimization time that scales up with the number of target gains.…”

“…An alternative approach for Raman amplifier pump power optimization uses bio-inspired evolutionary strategies (ES) such as genetic algorithm [9]- [11], differential evolution (DE) algorithm [12] [13], and most recently co-variance matrix adaptation evolutionary strategy [14]. Starting from a large set of random and non-trivial solutions, these populationdriven approaches navigate through the solution space looking for the target performance.…”

“…This helps to speed up the optimization by allowing parallel evaluation of the solutions within the population. However, for the solution to be useful experimentally, a thorough physical layer characterization is also required [14]. This makes the offline ES approach also dependent on the experimental scenario.…”

Fiber-Agnostic Machine Learning-Based Raman Amplifier Models

“…Although highly accurate, these numerical methods are very complex to be embedded into real-time QoT applications. They also require a detailed physical layer parameters extraction, as in [31], to accurately predict the gain profile. Finally, their nondifferentiable characteristics limit their application to gradient-free system optimization approaches.…”

“…We experimentally demonstrate that such a general model based on numerical data and enhanced by TL has great generalization performance for different optical fibers. Moreover, it also avoids the need to embed a full numerical ODE solver into the QoT estimator as in [30], [31].…”