Skip to content

Parameter Search

ssqp::param_search (ssqpy.solvers.search in Python) provides online parameter identification for dynamics that expose tunable parameters (e.g. friction, mass, or other physical coefficients embedded in a custom StageDynamics), plus a small utility for searching over SSQP hyperparameters.

Both searchers take (state, control, resulting next-state) data points and use Armijo line search to update a dynamics model's internal parameters so its predictions better match observed transitions. This is useful for adapting a model online or from offline rollout data.

from ssqpy.solvers import search

searcher = search.DynamicsParamGradientDescent(dynamics, min_data_size=100)

updated = searcher.addData(x_prev, u, x_actual)  # returns whether parameters changed

  • DynamicsParamGradientDescent: accumulates data points and performs a gradient-descent update once at least min_data_size points have been collected (setMinimumDataSize to change this later).

  • DynamicsParamGaussNewton: maintains a sliding window of window_size data points and performs a Gauss-Newton update against them. setWindowSize(window_size, prune_last=False) resizes the window; when shrinking, prune_last=True drops the most recent points, while the default instead prunes the most similar points, keeping the window's data diverse.

Both share addData(x, u, x_next) and clearData(), and both take armijo_beta/armijo_c1 to control the Armijo line search used for the parameter update step.

HyperParamSearch searches over a set of candidate ssqpParams across a set of Model problems to find the configuration that performs best:

hp_search = search.HyperParamSearch(candidate_params_list, model_list)
best_params = hp_search.search()

or build it up incrementally with addHyperparams(params) / addProblem(model) before calling search().