Other answers present the pattern correctly, however I don't feel they are clear enough. Unfortunately the link I've provided does the same, so I'll try to demonstrate what's the Strategy's spirit, IMHO.
Main thing about strategy is to have a general procedure, with some of its details (behaviours) abstracted, allowing them to be changed transparently.
Consider an gradient descent optimization algorithm - basically, it consists of three actions:
- gradient estimation
- step
- objective function evaluation
Usually one chooses which of these steps they need abstracted and configurable. In this example it seems that evaluation of the objective function is not something that you can do in more than one way - you always just ... evaluate the function.
So, this introduces two different strategy (or policy) families then:
interface GradientStrategy
{
double[] CalculateGradient(Function objectiveFunction, double[] point);
}
interface StepStrategy
{
double[] Step(double[] gradient, double[] point);
}
where of course Function
is something like:
interface Function
{
double Evaluate(double[] point);
}
interface FunctionWithDerivative : Function
{
double[] EvaluateDerivative(double[] point);
}
Then, a solver using all these strategies would look like:
interface Solver
{
double[] Maximize(Function objectiveFunction);
}
class GradientDescentSolver : Solver
{
public Solver(GradientStrategy gs, StepStrategy ss)
{
this.gradientStrategy = gs;
this.stepStrategy = ss;
}
public double[] Maximize(Function objectiveFunction)
{
// choosing starting point could also be abstracted into a strategy
double[] currentPoint = ChooseStartingPoint(objectiveFunction);
double[] bestPoint = currentPoint;
double bestValue = objectiveFunction.Evaluate(bestPoint);
while (...) // termination condition could also
// be abstracted into a strategy
{
double[] gradient = this.gradientStrategy.CalculateGradient(
objectiveFunction,
currentPoint);
currentPoint = this.stepStrategy.Step(gradient, currentPoint);
double currentValue = objectiveFunction.Evaluate(currentPoint);
if (currentValue > bestValue)
{
bestValue = currentValue;
bestPoint = currentPoint;
}
else
{
// terminate or step back and reduce step size etc.
// this could also be abstracted into a strategy
}
}
return bestPoint;
}
private GradientStrategy gradientStrategy;
private StepStrategy stepStrategy;
}
So the main point is that you have some algorithm's outline, and you delegate particular, general steps of this algorithm to strategies or policies. Now you could implement GradientStrategy
which works only for FunctionWithDerivative
(casts down) and just uses function's analytical derivative to obtain the gradient. Or you could have another one implementing stochastic version of gradient estimation. Note, that the main solver does not need to know about how the gradient is being calculated, it just needs the gradient. The same thing goes for the StepStrategy
- it can be a typical step policy with single step-size:
class SimpleStepStrategy : StepStrategy
{
public SimpleStepStrategy(double stepSize)
{
this.stepSize = stepSize;
}
double[] Step(double[] gradient, double[] point)
{
double[] result = new double[point.Length];
for (int i = 0;i < result.Length;++i)
{
result[i] = point[i] + this.stepSize * gradient[i];
}
return result;
}
private double stepSize;
}
, or a complicated algorithm adjusting the step-size as it goes.
Also think about the behaviours noted in the comments in the code: TerminationStrategy
, DeteriorationPolicy
.
Names are just examples - they're probably not the best, but I hope they give the intent. Also, usually best to stick with one version (Strategy or Policy).