Solves the DDE defined by
prob using the algorithm
alg. If no algorithm is given, a default algorithm will be chosen.
The recommended method for DDE problems are the
MethodOfSteps algorithms. These are constructed from an OrdinaryDiffEq.jl algorithm as follows:
MethodOfSteps(alg;constrained=false, fixedpoint_abstol = nothing, fixedpoint_reltol = nothing, fixedpoint_norm = nothing, max_fixedpoint_iters = 10)
alg is an OrdinaryDiffEq.jl algorithm. Most algorithms should work.
The standard choice is
MethodOfSteps(OrwenZen5()). This is a highly efficient FSAL 5th order algorithm which optimizes its interpolation error and should handle most problems. For fast solving at where non-strict error control is needed, choosing
OrwenZen3() can do well. Using
BS3 is similar to the MATLAB
OrwenZen3() will have noticably less error for the same work. For algorithms where strict error control is needed, it is recommended that one uses
Stiff DDEs and Differential-Algebraic Delay Equations (DADEs)
For stiff DDEs, the SDIRK and Rosenbrock methods are very efficient as they will reuse the Jacobian in the unconstrained stepping iterations. One should choose from the methods which have stiff-aware interpolants for better stability.
MethodOfSteps(Rosenbrock23()) is a good low order method choice. Additionally, the
Rodas methods like
MethodOfSteps(Rodas4()) are good choices because of their higher order stiff-aware interpolant.
Additionally, DADEs can be solved by specifying the problem in mass matrix form. The Rosenbrock methods are good choices in these situations.
Lags are declared separately from their use. One can use any lag by simply using the interpolant of
h at that point. However, one should use caution in order to achieve the best accuracy. When lags are declared, the solvers can more efficiently be more accurate. Constant delays are propagated until the order is higher than the order of the integrator. If state-dependent delays are declared, the algorithm will detect discontinuities arising from these delays and adjust the step size such that these discontinuities are included in the mesh. This way, all discontinuities are treated exactly.
If there are undeclared lags, the discontinuities due to delays are not tracked. In this case, one should only use residual control methods like
RK4(), which is the current best choice, as these will step more accurately. Still, residual control is an error-prone method. We recommend setting the tolerances low (
1e-10) and only trusting the solution to a 2-3 decimal places of accuracy.
Special Keyword Arguments
minimal_solution- Allows the algorithm to delete past history when
save_everystepare false, and only constant lags are specified. Defaults to true. If lags can grow or some lags are undeclared this may need to be set to false since it might impact the quality of the solution otherwise.
initial_order- Order of discontinuity at the initial time point. Defaults to 0 if the evaluation of the history function at the initial time point does not equal the initial state, and 1 otherwise.
discontinuity_interp_points- Number of interpolation points used to track discontinuities arising from dependent delays. Defaults to 10. Only relevant if dependent delays are declared.
discontinuity_reltol- These are absolute and relative tolerances used by the check whether the time point at the beginning of the current step is a discontinuity arising from dependent delays. Defaults to 1/10^12 and 0. Only relevant if dependent delays are declared.
If the method is having trouble, one may want to adjust the parameters of the fixed-point iteration. Decreasing the absolute tolerance
fixedpoint_abstol and the relative tolerance
fixedpoint_reltol, and increasing the maximal number of iterations
max_fixedpoint_iters can help ensure that the steps are correct. If the problem still is not correctly converging, one should lower
dtmax. In the worst case scenario, one may need to set
constrained=true which will constrain timesteps to at most the size of the minimal lag and hence forces more stability at the cost of smaller timesteps.