Lecture 15: The Strategy and Decorator Patterns
Objectives of the lecture
This lecture introduces two design patterns: strategy and decorators. Both patterns are useful to enhance existing operations and implement combinations of existing implementations of operations.
The strategy and decorator patterns are very similar idioms of using delegation to refactor aspects of a class’s behavior into separate classes, for easier enhancement or modification, and tighter cohesiveness. The strategy pattern is somewhat more intuitive to motivate with real-world examples, so we’ll start there.
1 The strategy pattern
Let’s consider a straightforward game for which the technique for winning is easily, thoroughly mastered: TicTacToe. How might we play this game? Here are some ideas:
Play in an open square
Play to an open corner
Block the opponent from winning
Place a third piece in a row to win
All but the first of these aren’t complete strategies. If the corners are all full, for instance, then the second approach doesn’t have any recommendation to make. So perhaps more accurately, our complete strategies might be:
Play in an open square
Play to an open corner, or else play in an open square
Block the opponent from winning if they’re about to, otherwise play to an open corner, otherwise just play to an open square
Place a third piece in a row to win, or else block the opponent if they’re about to win, or else play in an open corner, or else just play to an open square
It’s probably intuitively clear to you, as an expert human player of TicTacToe,
that the final strategy above is pretty good.1It’s not perfect —
1.1 A strategy is just a function object
Suppose in our program we had a TicTacToeModel
containing the board
state, and a Piece
enumeration naming X
and O
. Our program
also contains two Player
objects (one for each Piece
), which may
call the model’s playMove(Piece player, int row, int col)
method to make
a move. How should the Player
s choose where to move? What information
would be needed to make an informed choice? We’d need to know the state of the
board, and also know which player we’re trying to choose a move for:
class Coord {
int row;
int col
}
interface TicTacToeStrategy {
Coord chooseMove(TicTacToeModel model, Piece forWhom);
}
Implementing our first strategy is pretty easy. It doesn’t care who it’s choosing a move for; it just seeks out the first open square:
class AnyOpenSquare implements TicTacToeStrategy {
public Coord chooseMove(TicTacToeModel model, Piece forWhom) {
for (int r = 0; r < 3; r++)
for (int c = 0; c < 3; c++)
if (model.getPieceAt(r, c) == null)
return new Coord(r, c);
return null;
}
}
(Note: we are returning null
at the moment to indicate the lack
of an available move. This not a great idea – might consider throwing
an exception instead, to indicate that our strategy could not succeed. But see
Avoiding null
: Multiple strategy signatures below for a better approach still.)
How might we implement the second strategy? We could implement the entirety of the logic (choose corners or else choose random), but part of that logic has been implemented already... Let’s implement just the new part, first:
class AnyOpenCorner implements TicTacToeStrategy {
public Coord chooseMove(TicTacToeModel model, Piece forWhom) {
if (model.getPieceAt(0, 0) == null) return new Coord(0, 0);
else if (model.getPieceAt(0, 2) == null) return new Coord(0, 2);
else if (model.getPieceAt(2, 0) == null) return new Coord(2, 0);
else if (model.getPieceAt(2, 2) == null) return new Coord(2, 2);
else ????????
}
}
How might we fill in the question-marks? This strategy looks for an open
corner, but if it can’t find one, it needs to use another approach to pick a
response. Fortunately, we have such an approach already: we could reuse
AnyOpenSquare
!
class AnyOpenCorner implements TicTacToeStrategy {
public Coord chooseMove(TicTacToeModel model, Piece forWhom) {
if (model.getPieceAt(0, 0) == null) return new Coord(0, 0);
else if (model.getPieceAt(0, 2) == null) return new Coord(0, 2);
else if (model.getPieceAt(2, 0) == null) return new Coord(2, 0);
else if (model.getPieceAt(2, 2) == null) return new Coord(2, 2);
else
return new AnyOpenSquare().chooseMove(model, forWhom);
}
}
This resolves the question-marks, at the cost of hardcoding which fallback mechanism to choose. But we can definitely be more flexible.
1.2 Strategies can be composed
The any-open-square strategy and the any-open-corner strategy are both functions that take in a board state and return a coordinate to move. We can easily, and generally, combine the two of them into a higher-order strategy that first tries one and then, if it fails, tries the other:
class TryTwo implements TicTacToeStrategy {
TicTacToeStrategy first, second;
public Coord chooseMove(TicTacToeModel model, Piece forWhom) {
Coord ans = this.first.chooseMove(model, forWhom);
if (ans != null) return ans; // the first strategy succeeded
return this.second.chooseMove(model, forWhom);
}
}
If we fill in the question-marks in AnyOpenCorner
with return
null
, as we effectively did with AnyOpenSquare
, then our full second
strategy is simply new TryTwo(new AnyOpenCorner(), new AnyOpenSquare())
.
We can continue implementing the other strategy components (blocking an
opponent, or going for an immediate win), and combining them with
TryTwo
. (These other components need the forWhom
argument —
More broadly, strategies can be composed in lots of different ways, analogous
to mapping, and-map, or-map, or other higher-order combinations. We can
generalize from TryTwo
to TryMany
(that takes a list of
strategies); we can generalize to randomly selecting among several strategies,
or more sophisticated choices among several strategies. We might possibly
try a BestOfThree
strategy that take in three strategies, try
them all, and pick the value that at least two strategies agreed on. The
possibilities here are vast.
1.3 Strategies can be dynamically selected
Suppose we wanted to build a TicTacToe game for players of varying ability, and wanted “easy”, “medium”, and “hard” difficulty levels. We could easily mix and match compositions among the four simple strategies above, and dynamically assign some such object at runtime:
User computerUser = ...;
Strategy easy = new TryTwo(new PlayToWin(), new AnyOpenSquare());
Strategy medium = new TryTwo(new PlayToWin(), new TryTwo(new DontLose(), new AnyOpenSquare()));
Strategy hard = new TryTwo(new PlayToWin(),
new TryTwo(new DontLose(),
new TryTwo(new AnyOpenCorner(), new AnyOpenSquare())));
if (difficulty == EASY)
computerUser.playStrategy(easy);
else if (difficulty == MEDIUM)
computerUser.playStrategy(medium);
else
computerUser.playStrategy(hard);
Rather than hardcoding the difficulty level of the game, we have refactored it out and made it easily configurable.
We can also define an AskUser
strategy...and now we have something the
controller can use to interact with a human player, in the same framework as
these other strategy choices.
(Note that in this scenario, the composition of strategies could be rearranged:
new TryTwo(new TryTwo(new PlayToWin(), new DontLose()),
new TryTwo(new PlayToCorner(), new AnyOpenSquare()));
behaves exactly the same as our hard
strategy above, even though the
component strategies are arranged in a tree instead of a list. Depending on
which combinations you pick, there may easily be many ways to express the
desired result.)
1.4 Strategies make for easy testing
If we supply two strategies to our controller and say “play the game with two
players using these respective strategies”, what could happen? If both
strategies are fully automatic, then we’ve built a great test harness to run
through the game automatically! Combining this with the Appendable
output we’ve seen in earlier lectures, and we have a straightforward means to
simulate a playthrough of the game. Moreover, testing different strategies
against each other lets us not only test whether they work properly, but also
how well they do at playing the game.
2 Avoiding null
: Multiple strategy signatures
Above, we built four partial strategies, that worked when possible and
returned null
otherwise. This is a bad idea, since the user of
these strategies is forced to check for nullness before using the result. But
it’s a mistake of our own making: we chose a return type for our strategy that
didn’t include the possibility of failure. A complete strategy in this
game will always be able to choose a place to play (assuming any open places
remain). But partial strategies might not always be able to do so.
Accordingly, since these are now two different purpose statements ("Find me a place!" versus "Try to find a place if you can"), we might encode them as two different strategy interfaces:
// An interface describing strategies whose return value cannot fail:
// they will always return a non-null Coord, or else
// throw an excpetion if they're called on a game that cannot have a move
interface InfallibleTicTacToeStrategy {
Coord chooseMove(TicTacToeModel model, Piece forWhom) throws IllegalStateException;
}
// An interface describing incomplete or partial strategies, that
// might successfully choose a move and might not.
interface FallibleTicTacToeStrategy {
Optional<Coord> chooseMove(TicTacToeModel model, Piece forWhom);
}
Then our partial strategies above would be better written as
class AnyOpenSquare implements FallibleTicTacToeStrategy {
public Optional<Coord> chooseMove(TicTacToeModel model, Piece forWhom) {
for (int r = 0; r < 3; r++)
for (int c = 0; c < 3; c++)
if (model.getPieceAt(r, c) == null)
return Optional.of(new Coord(r, c));
return Optional.empty();
}
}
class AnyOpenCorner implements FallibleTicTacToeStrategy {
public Optional<Coord> chooseMove(TicTacToeModel model, Piece forWhom) {
if (model.getPieceAt(0, 0) == null) return Optional.of(new Coord(0, 0));
else if (model.getPieceAt(0, 2) == null) return Optional.of(new Coord(0, 2));
else if (model.getPieceAt(2, 0) == null) return Optional.of(new Coord(2, 0));
else if (model.getPieceAt(2, 2) == null) return Optional.of(new Coord(2, 2));
else return Optional.empty();
}
}
class TryTwo implements FallibleTicTacToeStrategy {
FallibleTicTacToeStrategy first, second;
public Optional<Coord> chooseMove(TicTacToeModel model, Piece forWhom) {
Optional<Coord> ans = this.first.chooseMove(model, forWhom);
if (ans.isPresent()) return ans; // the first strategy succeeded
return this.second.chooseMove(model, forWhom);
}
}
Ultimately, our User
wants a complete strategy, not one that can
fail. So we need a way to convert our fallible strategies into an infallible
one:
class CompleteStrategyFromFallible implements InfallibleTicTacToeStrategy {
FallibleTicTacToeStrategy strategyToTry;
public Coord chooseMove(TicTacToeModel model, Piece forWhom) throws IllegalStateException {
Optional<Coord> maybeAns = this.strategyToTry.chooseMove(model, forWhom);
if (maybeAns.isPresent()) { return maybeAns.get(); }
throw new IllegalStateException("There are no possible moves chosen by this strategy!");
}
}
3 The decorator pattern
The decorator pattern is similar to the strategy pattern in terms of how
objects delegate from one to another, but the purpose is different. The
canonical example of decorators is a UI widget library, such as Swing. We
might have a basic JPanel
class that just describes a rectangle of
on-screen content. We might have a JScrollPanel
class, which is a
subclass of JPanel
, that visually wraps around some other JPanel
and adds scrollbars and the ability to scroll the view in one or more
directions. Or a JSplitPanel
, which surrounds two panels and produces a
split-screen effect. Or a JGridPanel
, or others... Each of these
classes obeys a fairly sophisticated interface (all the methods of the base
JPanel
class), and then does two things: it provides some new
functionality of its own, and delegates the rest of the functionality to inner
panels. Visually, each surrounding panel decorates the inner ones by
adding borders, scrollbars, splitters, etc. Just as with the strategy pattern,
each class is responsible for one fragment of functionality: the complete
functionality comes about by composing several decorators together around some
base object.
Unlike strategies, where there can often be useful strategies
that do not need to delegate at all; decorators are all about the delegation.
Also, unlike strategies, a decorator can never be a "partial" function —
4 Strategies, Decorators and Inheritance
When should you use a strategy pattern or a decorator pattern, instead of just
using inheritance to customize behavior? After all, don’t subclasses
specialize the behavior of their superclasses? For example, could we produce
the same effect by having our User
class just have an abstract method
chooseMove
, and then creating subclasses that define various
possible implementations for that method?
Yes...but only sometimes. The power of the strategy and decorator patterns comes from the dynamism inherent in ability to delegate from one simple strategy to the next, or from one decorator to its contained content. Rather than being fixed at compile time, we can use higher-order strategies to mix-and-match strategy pieces, or decorate panels into elaborate UIs, without having to hard-code those choices in advance. Additionally, the strategy or decorator classes are appealingly tiny, self-contained, and easy to read: they have very high cohesion, because they’re built to do exactly one thing and nothing else.
1It’s not perfect —