LISCO
A learned iterative solver for constrained optimization that predicts primal-dual step sizes to progressively approach KKT optimality conditions. It employs a self-supervised training scheme requiring no precomputed optimizer solutions as labels, and is designed to accelerate solving nonlinear model predictive control problems.
Timeline (last 90 days)
LISCO (Learned Iterative Solver for Constrained Optimization) learns to predict primal-dual step sizes to iteratively approach a solution satisfying KKT optimality conditions, rather than directly predicting the final control output.
LISCO adopts a self-supervised training approach, using the degree to which KKT conditions are satisfied as the objective, without requiring pre-computed optimizer solutions as labels.
The original LISCO adopts a dense monolithic network architecture whose network size grows linearly with the prediction horizon, making it difficult to scale to long-horizon, large-scale problems.
The temporal decomposition modification of the new architecture does not change the training loss function or the iterative logic of the solver, and is implemented solely at the architectural level.
The new solver can handle problems of larger scale than monolithic LISCO, demonstrating that temporal decomposition removes the horizon bottleneck.
Current validation remains limited to specific benchmark systems, and its generalization capability in higher dimensions, stronger nonlinearities, and real hardware deployment remains to be examined in future research.
All Facts (6)
LISCO (Learned Iterative Solver for Constrained Optimization) learns to predict primal-dual step sizes to iteratively approach a solution satisfying KKT optimality conditions, rather than directly predicting the final control output.
50%UnverifiedLISCO adopts a self-supervised training approach, using the degree to which KKT conditions are satisfied as the objective, without requiring pre-computed optimizer solutions as labels.
50%UnverifiedThe original LISCO adopts a dense monolithic network architecture whose network size grows linearly with the prediction horizon, making it difficult to scale to long-horizon, large-scale problems.
50%UnverifiedThe temporal decomposition modification of the new architecture does not change the training loss function or the iterative logic of the solver, and is implemented solely at the architectural level.
50%UnverifiedThe new solver can handle problems of larger scale than monolithic LISCO, demonstrating that temporal decomposition removes the horizon bottleneck.
50%UnverifiedCurrent validation remains limited to specific benchmark systems, and its generalization capability in higher dimensions, stronger nonlinearities, and real hardware deployment remains to be examined in future research.
50%