Penalty sequential convex optimisation with a trust region, the outer loop
of TrajOpt (Schulman et al. 2014, "Motion planning with sequential convex
optimization and convex collision checking", IJRR 33:1251, Algorithm 1)
over any iterate: the caller convexifies its problem about an iterate at a
penalty and solves the convex model within a region; this loop takes or
rejects the step by the ratio of the merit's actual to its predicted
decrease, moves the region by the caller's schedule (ISchedule), and
raises the penalty along the caller's ladder while the constraints are not
met, or steers it at every iterate (IOptions.steer). The merit is the objective plus the penalty times the constraints'
summed violation, the l1 exact penalty (Nocedal and Wright 2006,
"Numerical Optimization", 2nd ed., section 17.2). A problem still
infeasible at the last penalty is left to the caller to refuse by its own
reading.
Penalty sequential convex optimisation with a trust region, the outer loop of TrajOpt (Schulman et al. 2014, "Motion planning with sequential convex optimization and convex collision checking", IJRR 33:1251, Algorithm 1) over any iterate: the caller convexifies its problem about an iterate at a penalty and solves the convex model within a region; this loop takes or rejects the step by the ratio of the merit's actual to its predicted decrease, moves the region by the caller's schedule (
ISchedule), and raises the penalty along the caller's ladder while the constraints are not met, or steers it at every iterate (IOptions.steer). The merit is the objective plus the penalty times the constraints' summed violation, the l1 exact penalty (Nocedal and Wright 2006, "Numerical Optimization", 2nd ed., section 17.2). A problem still infeasible at the last penalty is left to the caller to refuse by its own reading.