On the analysis and emulation of nonlinear component characteristics

Bayesian Optimization for Nonlinear System Identification and Pre-Distortion in Cognitive Transmitters

“…According to the results shown in Fig. 4, Bayesian-based SI is able to reach the minimum identification error SI , 46% faster compared to the benchmarked approach [10]. This indicates that using BO to identify the optimal memory tap distribution of a Volterra filter brings the advantage of reducing the convergence time, what makes it more computationally suitable for CT.…”

“…In both works, the authors demonstrate the potential of using ML optimization tools, such as BO, to automate the search for optimal designs of neural networks that can compensate for transceiver and fiber noise. This triggers the potential of using BO to enhance the design of Volterra-based DPD filters, thus tackling scalability pitfalls and alternatively replacing conventional pruning/truncation techniques such as regularization [34] or grid-based heuristic methods [10]. Therefore, this work proposes a scheme that uses traditional Volterra series for DPD assisted by a modern ML-based optimization tool, i.e., BO.…”

“…Recently, a SI approach was introduced in [10]. In this scheme, a heuristic was proposed to optimize the number of orders and memory taps of the Volterra series and follows the described method.…”

“…To benchmark Bayesian-based SI with [10], we synthesized a 5 th -order Volterra model that emulates the response of a SUT and is depicted by Fig. 3, for which is known and equal to = [39,5,7,3,1].…”

“…However, despite the effectiveness of Volterrabased DPD methods, as transmission becomes more dynamic and demanding (symbol rates ≥ 60 GBd coupled with advanced modulation formats), optimizing the total number of orders and memory taps of the Volterra filter without increasing the complexity of the model becomes a challenging task. Common estimation approaches use empirical tuning [7], requiring manual configuration of the memory tap distribution, or heuristic procedures inspired by grid-search [10], which are computationally expensive and, hence, inappropriate for cognitive applications. Optionally, one way to address this optimization problem is via modern Machine Learning (ML) algorithms.…”