Tools to ease the uniformization of stopping criteria in iterative solvers.

When a solver is called on an optimization model, four outcomes may happen:

- the approximate solution is obtained, the problem is considered solved
- the problem is declared unsolvable (unboundedness, infeasibility ...)
- the maximum available resources are not sufficient to compute the solution
- some algorithm dependent failure happens

This tool eases the first three items above. It defines a type

```
mutable struct GenericStopping <: AbstractStopping
problem :: Any # an arbitrary instance of a problem
meta :: AbstractStoppingMeta # contains the used parameters and stopping status
current_state :: AbstractState # Current information on the problem
main_stp :: Union{AbstractStopping, Nothing} # Stopping of the main problem, or nothing
listofstates :: Union{ListStates, Nothing} # History of states
user_specific_struct :: Any # User-specific structure
```

The StoppingMeta provides default tolerances, maximum resources, ... as well as (boolean) information on the result.

The GenericStopping (with GenericState) provides a complete structure to handle stopping criteria. Then, depending on the problem structure, you can specialize a new Stopping by redefining a State and some functions specific to your problem.

We provide some specialization of the GenericStopping for optimization:

- NLPStopping with NLPAtX as a specialized State: for non-linear programming (based on NLPModels);
- LAStopping with GenericState: for linear algebra problems.
- LS_Stopping with LSAtT as a specialized State: for 1d optimization;
- more to come...

The tool provides two main functions:

`start!(stp :: AbstractStopping)`

initializes the time and the tolerance at the starting point and check wether the initial guess is optimal.`stop!(stp :: AbstractStopping)`

checks optimality of the current guess as well as failure of the system (unboundedness for instance) and maximum resources (number of evaluations of functions, elapsed time ...)

Stopping uses the informations furnished by the State to evaluate its functions. Communication between the two can be done through the following functions:

`update_and_start!(stp :: AbstractStopping; kwargs...)`

updates the states with informations furnished as kwargs and then call start!.`update_and_stop!(stp :: AbstractStopping; kwargs...)`

updates the states with informations furnished as kwargs and then call stop!.`fill_in!(stp :: AbstractStopping, x :: Iterate)`

a function that fill in all the State with all the informations required to correctly evaluate the stopping functions. This can reveal useful, for instance, if the user do not trust the informations furnished by the algorithm in the State.`reinit!(stp :: AbstractStopping)`

reinitialize the entries of the Stopping to reuse for another call.

Consult the HowTo tutorial to learn more about the possibilities offered by Stopping.

You can also access other examples of algorithms in the test/examples folder, which for instance illustrate the strenght of Stopping with subproblems:

- Consult the OptimSolver tutorial for more on how to use Stopping with nested algorithms.
- Check the Benchmark tutorial to see how Stopping can combined with SolverBenchmark.jl.
- Stopping can be adapted to closed solvers via a buffer function as in Buffer tutorial for an instance with Ipopt via NLPModelsIpopt.

Install and test the Stopping package with the Julia package manager:

```
pkg> add Stopping
pkg> test Stopping
```

You can access the most up-to-date version of the Stopping package using:

```
pkg> add https://github.com/vepiteski/Stopping.jl
pkg> test Stopping
pkg> status Stopping
```

As an example, a naive version of the Newton method is provided here. First we import the packages:

```
using LinearAlgebra, NLPModels, Stopping
```

We consider a quadratic test function, and create an uncontrained quadratic optimization problem using NLPModels:

```
A = rand(5, 5); Q = A' * A;
f(x) = 0.5 * x' * Q * x
nlp = ADNLPModel(f, ones(5))
```

We now initialize the *NLPStopping*. First create a State.

```
nlp_at_x = NLPAtX(ones(5))
```

We use unconstrained_check as an optimality function

```
stop_nlp = NLPStopping(nlp, nlp_at_x, optimality_check = unconstrained_check)
```

Note that, since we used a default State, an alternative would have been:

```
stop_nlp = NLPStopping(nlp)
```

Now a basic version of Newton to illustrate how to use Stopping.

```
function newton(stp :: NLPStopping)
#Notations
pb = stp.pb; state = stp.current_state;
#Initialization
xt = state.x
#First, call start! to check optimality and set an initial configuration
#(start the time counter, set relative error ...)
OK = update_and_start!(stp, x = xt, gx = grad(pb, xt), Hx = hess(pb, xt))
while !OK
#Compute the Newton direction (state.Hx only has the lower triangular)
d = (state.Hx + state.Hx' - diagm(0 => diag(state.Hx))) \ (- state.gx)
#Update the iterate
xt = xt + d
#Update the State and call the Stopping with stop!
OK = update_and_stop!(stp, x = xt, gx = grad(pb, xt), Hx = hess(pb, xt))
end
return stp
end
```

Finally, we can call the algorithm with our Stopping:

```
stop_nlp = newton(stop_nlp)
```

and consult the Stopping to know what happened

```
#We can then ask stop_nlp the final status
@test :Optimal in status(stop_nlp, list = true)
#Explore the final values in stop_nlp.current_state
printstyled("Final solution is $(stop_nlp.current_state.x)", color = :green)
```

We reached optimality, and thanks to the Stopping structure this simple looking algorithm verified at each step of the algorithm:

- time limit has been respected;
- evaluations of the problem are not excessive;
- the problem is not unbounded (w.r.t. x and f(x));
- there is no NaN in x, f(x), g(x), H(x);
- the maximum number of iteration (call to stop!) is limited.

Stopping is aimed as a tool for improving the reusability and robustness in the implementation of iterative algorithms. We warmly welcome any feedback or comment leading to potential improvements.

Future work will address more sophisticated problems such as mixed-integer optimization problems, optimization with uncertainty. The list of suggested optimality functions will be enriched with state of the art conditions.

02/22/2017

26 days ago

201 commits