Skip to content

Added a proper expected model which includes the gaps + Goldstein step acceptance

Carlos Mastalli requested to merge cmastall/crocoddyl:topic/fddp-expectation-model-goldstein-acceptance into devel

In this PR there is mainly needed improvements in the SolverFDDP plus a couple of extra changes. Here I list them:

  1. Added an expectation model that considers the gaps
  2. Efficiently computations of the expectation model. For this it was needed to define an argument for firstCalc inside expectedImprovement.
  3. Added the Goldstein condition which it's needed for constraint satisfaction.
  4. Defined d1 and d2 as global variables in order to avoid to run expecteImprovement inside callback. This safes computation time.
  5. Added unit-test that compares the expected improvement against SolverKKT. It's tested for two scenarios: LQR problem with and without bias term.
  6. Updated the walk.log file after changes 1 and 3.
  7. Returned a list, instead of tuple, for the SolverKKT.expectedImprovement function. This is compatible with the other solvers.

Merge request reports