Tunable-bias based optical neural network for reinforcement learning in path planning.

Area of Science:

• Photonics and Artificial Intelligence
• Optical Computing and Neural Networks

Background:

• Mach-Zehnder interferometers (MZIs) are widely used in optical neural networks (ONNs) due to their integration, reconfiguration, and robustness.
• However, incorporating bias, a crucial element in traditional neural networks, into ONNs and studying its impact remains underexplored.

Purpose of the Study:

• To propose and investigate a tunable-bias optical neural network (TBONN) that enhances MZI utilization and network representational capacity.
• To systematically analyze the effect of optical biases on ONN performance.

Main Methods:

• Development of a TBONN architecture featuring a unitary matrix layer and tunable optical biases.
• Systematic study of the underlying mechanisms and characteristics of TBONN.
• Implementation of TBONN for a two-dimensional dataset classification task and development of an optical deep Q network (ODQN) for path planning.

Main Results:

• TBONN with two biases achieved 97.1% average prediction accuracy, a 5% improvement over TBONN with zero biases (92.1%).
• The proposed ODQN algorithm demonstrated competitive performance against conventional deep Q networks, with 2.5x and 4.5x speedups for 2D and 3D grid worlds, respectively.
• Demonstrated strong robustness and an imprecision elimination method using on-chip training.

Conclusions:

• Adding tunable optical biases significantly enhances the performance and representational capacity of ONNs.
• TBONN offers a promising approach for accelerating complex computational tasks, including path planning, with potential for further acceleration in more demanding applications.
• The developed TBONN exhibits strong robustness, making it suitable for practical implementation with on-chip training.