Keskeiset käsitteet
Model-based approach for training feedback controllers in highly nonlinear environments using DT-STL and dropout to handle long-horizon temporal tasks efficiently.
Tiivistelmä
The content introduces a model-based approach for training feedback controllers in nonlinear environments using DT-STL. It addresses challenges with long-horizon tasks by proposing a novel gradient approximation algorithm based on dropout. The methodology aims to improve training efficiency and scalability for complex spatio-temporal tasks.
Abstract:
- Introduces model-based approach for training feedback controllers.
- Uses DT-STL to handle specific task objectives and safety constraints.
- Proposes a novel gradient approximation algorithm based on dropout.
- Demonstrates efficacy on motion planning applications with complex tasks.
Introduction:
- Neural networks used for feedback control in nonlinear environments.
- Challenges with optimizing cost functions for system behavior.
- Importance of spatio-temporal task objectives expressed in DT-STL.
Training Neural Network Control Policies:
- Utilizes recurrent neural networks for control synthesis.
- Challenges with vanishing and exploding gradients in long-horizon tasks.
- Introduces sampling-based gradient approximation inspired by dropout.
Extension to Long Horizon Temporal Tasks & Higher Dimensional Systems:
- Addresses challenges with critical predicates in control synthesis.
- Proposes safe re-smoothing technique to handle non-differentiable local maxima.
Computing the Sampled Gradient:
- Differentiates between original and sampled gradients for efficient computation.
- Illustrates the methodology through examples of trajectory sampling.
Lainaukset
"In each iteration, we pick some recurrent units to be 'frozen', effectively approximating the gradient propagation."
"Our key idea is to approximate the gradient during back-propagation by an approximation scheme similar to drop-out layers."