NEW! Master Swift design patterns with my latest book! >>

< Previous: Implementing GKGameModel: gameModelUpdates(for:) and apply()   Next: Wrap up >

Creating a GameplayKit AI using GKMinmaxStrategist

If you've made it this far then you have built a Four in a Row game where two players can place chips in the game slots and either win or draw, and you've also prepared your model data to be run through the new GameplayKit AI routines. But we haven't created the AI just yet: we've just added some methods to our game models to enable an AI to make choices.

In this final step, we're going to use a new class called GKMinmaxStrategist, which is a gameplay strategy that tries to MINimize losses while MAXimizing gains – hence the name minmax, or minimax. When you create a GKMinmaxStrategist you tell it how many moves it should look ahead, and also what it should do to break ties, i.e. if it has two or more moves that are equally good.

Once you've created the strategist object, you need to provide it a game model to examine (that's our Board class), then ask it either to make the best move or make a random good move. If you ask for the best move, you'll get given back a GKGameModelUpdate object (that's a Move in our game) that represents the best move. If you ask for a random good move you'll need to tell it how many it should consider good (i.e., pick one from the top 5), and you'll get back a random GKGameModelUpdate from that list of good moves.

Now, one thing to be aware of up front: running AI takes a long time, particularly if you have a high look ahead depth. As a result, you should run the AI on a background thread so that your user interface doesn't lock up, and only push work back to the main thread when you have a move ready to make.

Let's go ahead and implement GKMinmaxStrategist now. Open ViewController.swift in your editor, then import GameplayKit. Now add this property to the ViewController class:

var strategist: GKMinmaxStrategist!

One strategist is capable of handling more than one game (i.e., if the player restarts the game) just by changing its game model, so we only need to create one GKMinmaxStrategist object. As it's needed straight away, we might as well put this into viewDidLoad() – anywhere before the call to resetBoard() is fine:

strategist = GKMinmaxStrategist()
strategist.maxLookAheadDepth = 7
strategist.randomSource = nil

Having a maxLookAheadDepth of 7 is a significant amount of work, because of those look aheads is one move being made by the player or AI – and each of those moves can be in any of seven columns. If you intend to alter this number upwards, be prepared for exponentially slower processing.

The randomSource property of GKMinmaxStrategist is there as a tie-breaker: if two moves result in the same advantage for the AI, which should it take? Setting it to nil as above means "just return the first best move," but if you wanted to have the AI take a random best move then you could try something like this:

strategist.randomSource = GKARC4RandomSource()

Now that the strategist is created, it wants to be fed some data. This is done by setting its gameModel property to an object that conforms to the GKGameModel protocol – which by now you should immediately recognize as our Board class. So, whenever we reset the board, we need to feed the new board into the strategist so it stands ready to look for moves.

We've done all the hard work to prepare for this, so all you need to do is change the start of your resetBoard() method to the following:

func resetBoard() {
    board = Board()
    strategist.gameModel = board


At this point, the AI understands the state of play, and stands ready to look for good moves. With GKMinmaxStrategist this is done using the bestMove(for:) method, which accepts a GKModelPlayer as its parameter and returns a GKModelUpdate for the best move if it finds one.

Remember, however, that AI can take a long time to consider all options depending on the look ahead depth you specify, so we're going to wrap this call up in a new method: columnForAIMove(). This will return an optional integer: either the best column for a move, or nil to mean "no move found." We'll call this on a background thread so it can take as long as it needs.

Here's the code for columnForAIMove():

func columnForAIMove() -> Int? {
    if let aiMove = strategist.bestMove(for: board.currentPlayer) as? Move {
        return aiMove.column

    return nil

Once the AI has found a good move, we want to run that that move on the main thread, because it will involve user interface changes. I've wrapped this up in another new method called makeAIMove(in:): this takes the column to move on, then makes it happen. This method will find the next available slot for the selected column, then use add(chip:) to make the move on the model, and addChip(inColumn:) to make the move in the view.

Once the AI move has been made, we'll call continueGame() to check for a win or draw, then flip turns so the player is in control.

Here's the code for makeAIMove(in:):

func makeAIMove(in column: Int) {
    if let row = board.nextEmptySlot(in: column) {
        board.add(chip: board.currentPlayer.chip, in: column)
        addChip(inColumn: column, row:row, color: board.currentPlayer.color)


At this point our game is almost finished, but we still need to call those methods on the appropriate threads. All this will be done in one big method called startAIMove(), which is going to do a number of things:

  1. Dispatch some work to the background thread.
  2. Get the current time, then run columnForAIMove().
  3. Get the time again, compare the difference, and subtract that value from 1 second to form a delay value.
  4. Run makeAIMove(in:) on the main thread after that delay, to execute the move.

The delay is there so that the AI always waits at least one second before making its move, otherwise it might confuse the user. If the AI takes half a second to find a move, we subtract that from our one second minimum to wait for a further half a second, equalling one second in total from before starting the AI to executing the move.

Here's the first draft of startAIMove():

func startAIMove() { { [unowned self] in
        let strategistTime = CFAbsoluteTimeGetCurrent()
        guard let column = self.columnForAIMove() else { return }
        let delta = CFAbsoluteTimeGetCurrent() - strategistTime

        let aiTimeCeiling = 1.0
        let delay = aiTimeCeiling - delta

        DispatchQueue.main.asyncAfter(deadline: .now() + delay) {
            self.makeAIMove(in: column)

Now only one thing more is required to finish the game: we need to call startAIMove() when it's black's turn. Change your updateUI() method to this:

func updateUI() {
    title = "\('s Turn"

    if board.currentPlayer.chip == .black {

Now, the game works, and you could even ship it today if you really wanted, but before you hit Run I'd like to suggest two tiny changes that make the whole experience better.

First, what happens if a user starts tapping buttons while the AI is "thinking"? Well, the answer is "bad things" – our game lets them play as black, and gets confused very quickly. So, our first change will be to disable all the column buttons when the AI's move starts, then re-enable them when it's finished.

Second, if the AI takes a long time, how does the user know the app hasn't locked up? There's no indication the AI is thinking, but it's easy enough to add by showing a custom UIBarButtonItem containing a UIActivityIndicatorView. It's not much, but it's enough to show the app is alive and well.

We're going to make both of these changes at once. All the AI code lives in startAIMove() so we can disable the column buttons and show the thinking spinner in there too. Add these lines to the start of startAIMove(), before the call to async():

columnButtons.forEach { $0.isEnabled = false }

let spinner = UIActivityIndicatorView(activityIndicatorStyle: .gray)

navigationItem.leftBarButtonItem = UIBarButtonItem(customView: spinner)

If you haven't seen forEach before, it's a way of quickly looping through an array, executing some code on every item in that array. In our case, the $0 means "each button in the loop", and in this way all the buttons get disabled.

Once the AI has finished their move, makeAIMove(in:) will be called on the main thread, and that's our chance to undo these changes: we need to re-enable the column buttons, then destroy the thinking spinner. Add these two lines of code at the start of makeAIMove(in:):

columnButtons.forEach { $0.isEnabled = true }
navigationItem.leftBarButtonItem = nil

These tiny changes stop users from accidentally screwing things up, and also stop them from worrying your app has got stuck in a loop somewhere. It's polish, yes, but polish is frequently what separates good games from great games.

That's it: the app is done! You can run it now and see how quickly you can beat the AI. It ought not to be too hard – our heuristic isn't very good, so sometimes the AI will miss obvious moves, just like a real player.

Upgrade to the premium experience

Get all 40 projects in PDF and ePub, plus exclusive content that will take your Swift learning to the next level – buy the Hacking with Swift book today!

< Previous: Implementing GKGameModel: gameModelUpdates(for:) and apply()   Next: Wrap up >
Buy Practical iOS 12 Buy Pro Swift Buy Swift Design Patterns Buy Practical iOS 11 Buy Swift Coding Challenges Buy Server-Side Swift (Vapor Edition) Buy Server-Side Swift (Kitura Edition) Buy Hacking with macOS Buy Advanced iOS Volume One Buy Hacking with watchOS Buy Hacking with tvOS Buy Hacking with Swift Buy Dive Into SpriteKit Buy Swift in Sixty Seconds Buy Objective-C for Swift Developers Buy Beyond Code

Was this page useful? Let me know!

Click here to visit the Hacking with Swift store >>