Handle quadratic-plus-linear error terms
Closed this issue · 2 comments
Extending the optimizer so that it handles problems of the form:
||f(x)||^2 + a^T g(x)
where a is a known vector, and f(x) and g(x) do not necessarily have the same dimensionality.
(only copied from https://github.com/ethz-asl/planning_2d/pull/30/files#r29438688):
We could just introduce a new error term role, which can be added alternatively to an OptimizationProblem: a scalar error, that is not squared. They could very well be just of the same type as the current error terms. When added as scalar error the dimension could be asserted to be one.
An optimizer can then check (when it gets the problem attached) whether it supports that problem. The current would refuse problems including scalars.
An potential new RProp or other gradient based one would probably refuse in the beginning all non scalar errors or just square them.
Addressed