Multi-wavelength optical information processing with deep reinforcement learning.

Area of Science:

• Optoelectronics
• Machine Learning
• Optical Signal Processing

Background:

• Multi-wavelength optical information processing is crucial for optical neural networks and broadband signal processing.
• Frequency-selective responses, arising from fabrication, transmission, and environmental factors, degrade system performance.

Purpose of the Study:

• To introduce a novel deep reinforcement learning calibration (DRC) method for mitigating frequency-selective responses in multi-wavelength optical systems.
• To demonstrate the DRC method's ability to autonomously learn and adapt calibration strategies.

Main Methods:

• The study employs a deep reinforcement learning calibration (DRC) approach, drawing inspiration from the deep deterministic policy gradient training strategy.
• The DRC method facilitates continuous, autonomous learning from the system to accumulate calibration knowledge.
• The method was tested on systems utilizing dispersion compensating fiber, micro-ring resonator array, and Mach-Zehnder interferometer array with multi-wavelength optical carriers.

Main Results:

• The DRC method successfully completed signal processing functions within 21 iterations across tested systems.
• The approach demonstrated superior adaptability and learning capabilities compared to traditional calibration methods.
• The system achieved efficient and accurate control for complex optical tasks.

Conclusions:

• The deep reinforcement learning calibration (DRC) method offers an effective solution for frequency-selective responses in multi-wavelength optical systems.
• This autonomous learning approach provides robust and adaptable calibration for advanced optical applications.
• The DRC method is suitable for accelerating optical convolution, microwave photonic signal processing, and optical network routing.