Termination Watcher

This topic contains the following sections:

Implementation
Code sample
See Also

Termination watcher is a utility class for the optimization algorithm stop conditions customization. This class just allows user to simplify his code.

The termination watcher allows user control next parameters:

Total iterations count;
Total execution time;
Objective function best value improvement.

Implementation

Next most important methods and properties are featured in the class:

Method	Description
TerminationWatcher(NullableDouble, NullableInt32, NullableTimeSpan, NullableInt64)	The TerminationWatcher constructor that allows to setup termination conditions.
Update(Double, Object)	Updates the TerminationWatcher state and properties. The NaN values are ignored without errors but the EvaluationsLimit and Timeout are checked.

Termination conditions properties:

Property	Description
EvaluationsLimit	Maximum number of objective function evaluations till termination.
Timeout	The algorithm execution timeout till termination flag.
Epsilon	Minimal fluctuation of the best objective value for next point evaluation.
EpsilonInterval	Best objective value fluctuation analysis interval.

Execution properties:

Property	Description
IsTerminalState	The terminal conditions have met flag.
EvaluationsCount	The total number of evaluation performed.
ElapsedTime	The total elapsed time from this object creation.

Objective function properties:

Property	Description
BestObjectiveValue	The best observed objective value.
BestObjectiveIndex	The best observed objective value index.
BestObjectiveModel	The optional model, associated with the best observed objective value.

Code sample

The next code generates test objective function and process it with the TerminationWatcher. The results of such processing are presented on the next figure.

TW Best Objective

Copy

 1using System;
 2using FinMath.MachineLearning;
 3using FinMath.Statistics;
 4
 5namespace FinMath.Samples
 6{
 7    class TerminationWatcherSample
 8    {
 9        private class RawRandomObjective
10        {
11            private RandomGenerator random = new RandomGenerator(42);
12            private long i = 0;
13
14            public double Next => Math.Exp(i++ / -100.0) + random.NextDouble() + Math.PI;
15        }
16
17        static void Main()
18        {
19            var objective = new RawRandomObjective();
20
21            var tw = new TerminationWatcher(1e-3, 1000, new TimeSpan(0, 1, 0));
22            while (!tw.IsTerminalState)
23            {
24                tw.Update(objective.Next);
25            }
26
27            Console.WriteLine($"Optimization stops on the {tw.EvaluationsCount} iteration in {tw.ElapsedTime} with best objective value {tw.BestObjectiveValue} at {tw.BestObjectiveIndex} step.");
28        }
29    }
30}

Other Resources

MachineLearning

NeuralNetwork