A simplified neural network solution through problem decomposition: the case of the truck backer-upper

Area of Science:

• Robotics and Artificial Intelligence
• Control Systems Engineering

Background:

• Previous research utilized large feedforward neural networks (25 hidden units) for tractor-trailer truck steering during reverse maneuvers, requiring extensive training.
• The complexity of these networks and training demands motivated the search for more efficient control solutions.

Purpose of the Study:

• To demonstrate a significantly simpler solution for the tractor-trailer truck backer-upper problem.
• To explore problem decomposition into subtasks for designing efficient neural network controllers.

Main Methods:

• Decomposing the truck backing control problem into manageable subtasks.
• Hard-wiring control laws derived from subtasks into a simplified neural network architecture.
• Comparing the performance of the simplified network against a larger, conventionally trained network.

Main Results:

• A simplified neural network controller with only two hidden units was developed.
• This simplified controller achieved performance comparable to the larger 25-hidden-unit network.
• The hard-wiring approach proved effective in creating an efficient control system.

Conclusions:

• A highly effective and simpler solution exists for the tractor-trailer truck backing problem.
• Decomposing complex control tasks into subtasks allows for the design of more efficient neural network controllers.
• This modular approach facilitates the construction of complex controllers from basic, pre-defined control laws.