Reinforcement Learning Neural Networks for Optical Communications.
Abstract
The objective of this work is to utilize neural networks to find new methods for optimizing high performance broadband fiber-optic communication links. In typical broadband analog optical communication links, the dominant distortion comes from the laser transmitter. The electrical-to-optical transfer characteristics of both electro-optic external modulators and semiconductor lasers are nonlinear and create both odd- and even-order harmonic distortions of the modulating signal. One cost-effective method to cancel device nonlinearities in direct modulated lasers is by electronic predistortion. For our previous work, based on the simulated annealing learning algorithm utilized for neural network learning, a novel algorithm was developed to obtain the initial parameters of predistortion and laser circuits, and it has been used to linearize the Distributed FeedBack (DFB) semiconductor laser transmitters. Because predistortion is not self-aligned for optimal performance, this type of laser transmitter would have to be readjusted in the field at various times. This readjustment is necessary in order to maintain optimal performance with variances in device performance due to drifting, aging, or possible changes in nonlinearities when the bias point of the laser changes, as derived from optical power feedback from a laser back facet monitor. The alternative to periodic hand-retuning the transmitter circuits is to use active techniques to monitor the transmitter performance and to compensate for the linearization from the measured performance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 02, 1996
- Accession Number
- ADA303321
Entities
People
- Michael Salour