Permalink
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
344 lines (272 sloc) 13.9 KB

Integrator Interface

The integrator interface gives one the ability to interactively step through the numerical solving of a differential equation. Through this interface, one can easily monitor results, modify the problem during a run, and dynamically continue solving as one sees fit.

Initialization and Stepping

To initialize an integrator, use the syntax:

integrator = init(prob,alg;kwargs...)

The keyword args which are accepted are the same Common Solver Options used by solve and the returned value is an integrator which satisfies typeof(integrator)<:DEIntegrator. One can manually choose to step via the step! command:

step!(integrator)

which will take one successful step. Additonally:

step!(integrator,dt[,stop_at_tdt=false])

passing a dt will make the integrator keep stepping until integrator.t+dt, and setting stop_at_tdt=true will add a tstop to force it to step to integrator.t+dt

To check whether or not the integration step was successful, you can call check_error(integrator) which returns one of the Return Codes (RetCodes).

This type also implements an iterator interface, so one can step n times (or to the last tstop) using the take iterator:

for i in take(integrator,n) end

One can loop to the end by using solve!(integrator) or using the iterator interface:

for i in integrator end

In addition, some helper iterators are provided to help monitor the solution. For example, the tuples iterator lets you view the values:

for (u,t) in tuples(integrator)
  @show u,t
end

and the intervals iterator lets you view the full interval:

for (tprev,uprev,u,t) in intervals(integrator)
  @show tprev,t
end

Additionally, you can make the iterator return specific time points via the TimeChoiceIterator:

ts = linspace(0,1,11)
for (u,t) in TimeChoiceIterator(integrator,ts)
  @show u,t
end

Lastly, one can dynamically control the "endpoint". The initialization simply makes prob.tspan[2] the last value of tstop, and many of the iterators are made to stop at the final tstop value. However, step! will always take a step, and one can dynamically add new values of tstops by modifiying the variable in the options field: add_tstop!(integrator,new_t).

Finally, to solve to the last tstop, call solve!(integrator). Doing init and then solve! is equivalent to solve.

DiffEqBase.step!
DiffEqBase.check_error
DiffEqBase.check_error!

Handing Integrators

The integrator<:DEIntegrator type holds all of the information for the intermediate solution of the differential equation. Useful fields are:

  • t - time of the proposed step
  • u - value at the proposed step
  • p - user-provided data
  • opts - common solver options
  • alg - the algorithm associated with the solution
  • f - the function being solved
  • sol - the current state of the solution
  • tprev - the last timepoint
  • uprev - the value at the last timepoint

The p is the data which is provided by the user as a keyword arg in init. opts holds all of the common solver options, and can be mutated to change the solver characteristics. For example, to modify the absolute tolerance for the future timesteps, one can do:

integrator.opts.abstol = 1e-9

The sol field holds the current solution. This current solution includes the interpolation function if available, and thus integrator.sol(t) lets one interpolate efficiently over the whole current solution. Additionally, a a "current interval interpolation function" is provided on the integrator type via integrator(t). This uses only the solver information from the interval [tprev,t] to compute the interpolation, and is allowed to extrapolate beyond that interval.

Note about mutating

Be cautious: one should not directly mutate the t and u fields of the integrator. Doing so will destroy the accuracy of the interpolator and can harm certain algorithms. Instead if one wants to introduce discontinuous changes, one should use the Event Handling and Callback Functions. Modifications within a callback affect! surrounded by saves provides an error-free handling of the discontinuity.

As low-level alternative to the callbacks, one can use set_t!, set_u! and set_ut! to mutate integrator states. Note that certain integrators may not have efficient ways to modify u and t. In such case, set_*! are as inefficient as reinit!.

DiffEqBase.set_t!
DiffEqBase.set_u!
DiffEqBase.set_ut!

Integrator vs Solution

The integrator and the solution have very different actions because they have very different meanings. The typeof(sol) <: DESolution type is a type with history: it stores all of the (requested) timepoints and interpolates/acts using the values closest in time. On the other hand, the typeof(integrator)<:DEIntegrator type is a local object. It only knows the times of the interval it currently spans, the current caches and values, and the current state of the solver (the current options, tolerances, etc.). These serve very different purposes:

  • The integrator's interpolation can extrapolate, both forward and backward in in time. This is used to estimate events and is internally used for predictions.
  • The integrator is fully mutable upon iteration. This means that every time an iterator affect is used, it will take timesteps from the current time. This means that first(integrator)!=first(integrator) since the integrator will step once to evaluate the left and then step once more (not backtracking). This allows the iterator to keep dynamically stepping, though one should note that it may violate some immutablity assumptions commonly made about iterators.

If one wants the solution object, then one can find it in integrator.sol.

Function Interface

In addition to the type interface, a function interface is provided which allows for safe modifications of the integrator type, and allows for uniform usage throughout the ecosystem (for packages/algorithms which implement the functions). The following functions make up the interface:

Saving Controls

  • savevalues!(integrator): Adds the current state to the sol.

Caches

  • get_tmp_cache(integrator): Returns a tuple of internal cache vectors which are safe to use as temporary arrays. This should be used for integrator interface and callbacks which need arrays to write into in order to be non-allocating. The length of the tuple is dependent on the method.
  • user_cache(integrator): Returns an iterator over the user-facing cache arrays.
  • u_cache(integrator): Returns an iterator over the cache arrays for u in the method. This can be used to change internal values as needed.
  • du_cache(integrator): Returns an iterator over the cache arrays for rate quantities the method. This can be used to change internal values as needed.
  • full_cache(integrator): Returns an iterator over the cache arrays of the method. This can be used to change internal values as needed.

Stepping Controls

  • u_modified!(integrator,bool): Bool which states whether a change to u occurred, allowing the solver to handle the discontinuity. By default, this is assumed to be true if a callback is used. This will result in the re-calculation of the derivative at t+dt, which is not necessary if the algorithm is FSAL and u does not experience a discontinuous change at the end of the interval. Thus if u is unmodified in a callback, a single call to the derivative calculation can be eliminated by u_modified!(integrator,false).
  • get_proposed_dt(integrator): Gets the proposed dt for the next timestep.
  • set_proposed_dt!(integrator,dt): Sets the proposed dt for the next timestep.
  • set_proposed_dt!(integrator,integrator2): Sets the timestepping of integrator to match that of integrator2. Note that due to PI control and step acceleration this is more than matching the factors in most cases.
  • proposed_dt(integrator): Returns the dt of the proposed step.
  • terminate!(integrator): Terminates the integrator by emptying tstops. This can be used in events and callbacks to immediately end the solution process.
  • change_t_via_interpolation!(integrator,t,modify_save_endpoint=Val{false}): This option lets one modify the current t and changes all of the corresponding values using the local interpolation. If the current solution has already been saved, one can provide the optional value modify_save_endpoint to also modify the endpoint of sol in the same manner.
  • add_tstop!(integrator,t): Adds a tstop at time t.
  • add_saveat!(integrator,t): Adds a saveat time point at t.

Resizing

  • resize!(integrator,k): Resizes the DE to a size k. This chops off the end of the array, or adds blank values at the end, depending on whether k>length(integrator.u).
  • resize_non_user_cache!(integrator,k): Resizes the non-user facing caches to be compatible with a DE of size k. This includes resizing Jacobian caches. Note that in many cases, resize! simple resizes user_cache variables and then calls this function. This finer control is required for some AbstractArray operations.
  • deleteat_non_user_cache!(integrator,idxs): deleteat!s the non-user facing caches at indices idxs. This includes resizing Jacobian caches. Note that in many cases, deleteat! simple deleteat!s user_cache variables and then calls this function. This finer control is required for some AbstractArray operations.
  • addat_non_user_cache!(integrator,idxs): addat!s the non-user facing caches at indices idxs. This includes resizing Jacobian caches. Note that in many cases, addat! simple addat!s user_cache variables and then calls this function. This finer control is required for some AbstractArray operations.
  • deleteat!(integrator,idxs): Shrinks the ODE by deleting the idxs components.
  • addat!(integrator,idxs): Grows the ODE by adding the idxs components. Must be contiguous indices.

Reinit

The reinit function lets you restart the integration at a new value. The full function is of the form:

reinit!(integrator::ODEIntegrator,u0 = integrator.sol.prob.u0;
  t0 = integrator.sol.prob.tspan[1], tf = integrator.sol.prob.tspan[2],
  erase_sol = true,
  tstops = integrator.opts.tstops_cache,
  saveat = integrator.opts.saveat_cache,
  d_discontinuities = integrator.opts.d_discontinuities_cache,
  reset_dt = (integrator.dtcache == zero(integrator.dt)) && integrator.opts.adaptive,
  reinit_callbacks = true, initialize_save = true,
  reinit_cache = true)

u0 is the value to start at. The starting time point and end point can be changed via t0 and tf. erase_sol allows one to start with no other values in the solution, or keep the previous solution. tstops, d_discontinuities, and saveat are reset as well, but can be ignored. reset_dt is a boolean for whether to reset the current value of dt using the automatic dt determination algorithm. reinit_callbacks is whether to run the callback initializations again (and initialize_save is for that). reinit_cache is whether to re-run the cache initialization function (i.e. resetting FSAL, not allocating vectors) which should usually be true for correctness.

Additionally, once can access auto_dt_reset!(integrator::ODEIntegrator) which will run the auto dt initialization algorithm.

Misc

  • get_du(integrator): Returns the derivative at t.
  • get_du!(out,integrator): Write the current derivative at t into out.
  • check_error(integrator): Checks error conditions and updates the retcode.

Note

Note that not all of these functions will be implemented for every algorithm. Some have hard limitations. For example, Sundials.jl cannot resize problems. When a function is not limited, an error will be thrown.

Additional Options

The following options can additionally be specified in init (or be mutated in the opts) for further control of the integrator:

  • advance_to_tstop: This makes step! continue to the next value in tstop.
  • stop_at_next_tstop: This forces the iterators to stop at the next value of tstop.

For example, if one wants to iterate but only stop at specific values, one can choose:

integrator = init(prob,Tsit5();dt=1//2^(4),tstops=[0.5],advance_to_tstop=true)
for (u,t) in tuples(integrator)
  @test t ∈ [0.5,1.0]
end

which will only enter the loop body at the values in tstops (here, prob.tspan[2]==1.0 and thus there are two values of tstops which are hit). Addtionally, one can solve! only to 0.5 via:

integrator = init(prob,Tsit5();dt=1//2^(4),tstops=[0.5])
integrator.opts.stop_at_next_tstop = true
solve!(integrator)

Plot Recipe

Like the DESolution type, a plot recipe is provided for the DEIntegrator type. Since the DEIntegrator type is a local state type on the current interval, plot(integrator) returns the solution on the current interval. The same options for the plot recipe are provided as for sol, meaning one can choose variables via the vars keyword argument, or change the plotdensity / turn on/off denseplot.

Additionally, since the integrator is an iterator, this can be used in the Plots.jl animate command to iteratively build an animation of the solution while solving the differential equation.

For an example of manually chaining together the iterator interface and plotting, one should try the following:

using DifferentialEquations, DiffEqProblemLibrary, Plots

# Linear ODE which starts at 0.5 and solves from t=0.0 to t=1.0
prob = ODEProblem((u,p,t)->1.01u,0.5,(0.0,1.0))

using Plots
integrator = init(prob,Tsit5();dt=1//2^(4),tstops=[0.5])
pyplot(show=true)
plot(integrator)
for i in integrator
  display(plot!(integrator,vars=(0,1),legend=false))
end
step!(integrator); plot!(integrator,vars=(0,1),legend=false)
savefig("iteratorplot.png")

Iterator Plot