BRUNO SOMMER: Hello everyone and welcome.
My name is Bruno Sommer, I'm a game technologies engineer here at Apple.
And today I'm very excited to be able to introduce you to GameplayKit.
Apple's first dedicated Gameplay framework.
We have a lot of solutions to the visual part of making game on our platform things like SpriteKit, SceneKit and Metal.
The gameplay is another really important part of that game development puzzle.
As it turns out, there is hard problems in the gameplay space, things like AI, pathfinding, autonomous movement.
We firmly believe that experience shouldn't be a barrier to prevent our developers making great and compelling games.
So going forward we want you guys to be able to focus on bringing your cool ideas to life.
And we'll do the heavy lifting on the back end to make that happen.
So our mission when we set out to make GameplayKit was very clear.
We wanted to make a simple, yet powerful solution API of gameplay solutions.
Now this is things like common design patterns and architectures so we can all start speaking the same gameplay language.
And there's also a number of standard gameplay algorithms that is applicable to a wide variety of game genres.
And it is also very important to us that this remains graphic and engine agnostic, so while GameplayKit is separate from a lot of those visual frameworks I talked about, it plays really nicely with them.
It plays nice with SpriteKit, SceneKit, Metal, and more.
So here we have GameplayKit and the seven major features that make it up.
And these are components which are a really great way to structure your game objects and game logic.
State machines, which describe the statefulness in our games and the various state changes of our game objects.
Agents, which are autonomously moving entities that are controlled by realistic behaviors and goals.
Pathfinding, which deals with navigation graph generation and how we move our entities between the passable areas in our game world.
We also have a great MinMax AI solution which is a really great way to give life to our computer-controlled opponents.
There is a number of game quality random sources and random distributions at your disposal.
And last we have a rule system, which are a really great way to model discreet and fuzzy logic.
There's a lot to cover today.
Let's go ahead and jump right in with entities and components.
I want to pose sort of this classic problem with inheriting from common game objects.
Here we have a tower defense game with a simple projectile tower and archer classes.
We have some shared functionality here.
We have shooting, and we have moving, and we have being targeted.
Let's take shooting for example.
We want towers and archers to both be able to shoot.
Where then do we put our shoot function?
One option might be to simply copy and paste it between the tower and archer classes, but now I have two spots in my code that share functionality, and if I ever want to update that functionality, there is now two spots where I need to update it.
And if I only update it in one I'm undoubtably going to get some really weird behavior.
So our only real option in this inheritance model I've described, is to move shared functionality higher in the tree.
So here we have a shoot function we might put it in the game object class or some common-based class.
Now the problem with this approach is that as we get more and more shared functionality in our games we're forced to move it higher and higher in the hierarchies.
And our basic game objects become anything but basic.
They become large and hard to understand, hard to maintain, and hard to collaborate on.
Let's take a look at how we solve this problem using entities and components.
You see here we still have our three objects: projectile, tower, archer.
But now instead of them having functionality in an inheritance sense, being a mover, being a shooter, or being targetable, they instead have these objects we call components which encapsulate singular elements of our game logic, so here we have a MoveComponent that deals with moving, a ShootComponent that deals with shooting, and a TargetComponent, what it means to be targetable.
So we gain these really nice little black boxes of singular functionality, that are loosely rather than tightly coupled with the hierarchy.
So we see now that entities and components are a great way to organize our game logic.
For one, they're easy to maintain because they're these nice black boxes of incapsulated functionality; they tend to be simpler.
We also have a really nice collaboration with the entities and components.
Now I can have one developer work on one component and another developer working on yet another component, and they don't necessarily need to know the intimate details between these components.
We also get nice scaling with complexity.
What I mean by that, in that class and inheritance model my hierarchy is grows wide and tall as my game gets more complex.
With entities and components it just grows wider in that width is no longer a detriment.
It's really a toolbox.
Any time I want to make a new entity in the game I simply look at the components I have available, choose the appropriate ones or perhaps implement a new one.
And with entities and components we get really easy access to dynamic behavior.
Let's think back to the tower defense example.
Perhaps I want to implement a magic spell that roots archers to the ground so they can no longer move.
One way to represent this might be to simply temporarily remove it's MoveComponent.
This implicitly tells the rest of my game that it can no longer move.
And I get the added benefit of the rest of my game not needing to know the intimate details of magic spells.
So let's go ahead and take a look at the classes.
Here we have GKEntity.
This is our entity base class, and it's really just a simple collection of components.
It has the functions to add and remove components dynamically, as my entities functionality undoubtedly changes.
Also let's me access existing components by unique class type.
When I call update on my GKEntity it's going to automatically update all of the components that it has.
So thinking back to that example, projectile, tower, and archer would all be GKEntities.
Here we have our GKComponent class.
Now you subclass this any time you want to add those singular bits of functionality to your game.
And you do that in a number of ways.
Properties on your components become state information about those components.
So you can imagine the ShootComponent here would likely have a damage property that describes how much damage it's projectiles do.
You also implement custom selectors that extend functionally and tell the rest of your game how to communicate with your component.
So the MoveComponent here for example would likely have a move to position function that you would call from the input or game controller code.
As I mentioned before components are automatically updated by their entity's update and you can optionally implement any time based logic in updateWithDeltaTime.
So undoubtedly a need will arise where you need finer control over the order or how your components update, and for that we're providing GKComponentSystem.
This is a collection of components from different entities, but they're all the same class type.
And you use this when update order is somehow intrinsically important to your game.
Perhaps I want to update AI after my movement code because I want my AI to deal with the most up to date position information available.
And it's important to note that the components that are placed in these component systems no longer update with their entities update.
It is up to you to call the component systems update at the correct time to update all these entities.
So thinking back to our example again, we probably have a move system which would in turn have all the move components in my game, and I can use that to synchronize the movement between my various entities.
So lastly we have a code example of what using entities and components in GameplayKit looks like.
You see at the top here I'm going to make my archer entity, and then I'm going to make the three components that make up being an archer: the MoveComponent, the ShootComponent, and the TargetComponent.
And add those to my archer.
Then I'm going to make that moveSystem we talked about, passing in the MoveComponent's class, indicating this component system only deals with MoveComponents.
Then I'm going to add my archer's MoveComponent to that moveSystem and I'm good to go.
This archer and moveSystem are ready for use in my game.
So that's entities and components.
So let's move on to state machines.
I'm going to start with another example here.
Let's imagine some game where the player is being chased by ghosts, and sometimes he gets to take a power-up and chase, and maybe defeat the ghosts instead.
Here's an example of what a state machine to control that ghost might look like.
You see here we have the four states that a ghost can ever be in, chase for when the ghost is chasing the player, flee for when the player is chasing the ghost, defeated for when the ghost gets caught and gets defeated, and respawn sometime after the ghost is defeated before it comes back to life.
Now it's important to note here that only some of these state transitions are valid.
You see I move between chase and flee interchangeably here.
This makes sense based on the game I just described, sometimes the ghost does the chasing and sometimes the player does the chasing.
And of course we only go to defeated from flee, this is the only time that the player can actually defeat the ghost, when he has that power-up and is chasing the ghost.
Then we go from respawn to defeated, this again makes sense and after we respawn we go right into chase.
This is our initial state.
After ghosts respawn, they go right back to chasing the player.
So why are state machines so important in game development?
Well for a lot of games they're the backbone of many gameplay elements.
A ton of our common gameplay elements are full of state, things like animation, AI, UI, levels.
Anyone who's tried to bring life to a humanoid character in a game is undoubtedly familiar with the state machine on the right.
We usually have an IdleAnimation, and a MoveAnimation, and an AttackAnimation, and move between them in meaningful ways.
So because this pattern is so pervasive in our code, we reimplemented it a lot, to what amounts to little more than boilerplate, and this take the form of really big switch statements or if else trees.
What if you could come up with some common implementation to remove that boilerplate, add a little bit of maintainability, and give us the benefit of being able to reuse our states and state machines throughout our game.
That's what we've done in GameplayKit.
So let's take a look at the classes.
Here we have GKStateMachine.
This is your general purpose finite state machine.
And what I mean by that is it's in one, and only one state at any given.
And it possesses all of the states that it can ever be in.
You call enterState on our state machines to cause those state transitions I was talking about.
And what happens under the hood, is it checks if that transition is valid, and if so, makes that change.
And it calls a number of call backs on the state objects.
We exit the state we were in, we enter the state we're going to, and we update the current state that the state machine is in.
So in that ghost example we'd probably have a GhostStateMachine.
It would in turn have those four states we were talking about.
Here we have our GKState abstract class.
And you implement your state based logic in a number of callbacks.
We give you an enter callback when the state is being entered, an exit callback when we're leaving the state, and an update callback when this is the current state in the state machine.
As I mentioned before, they're automatically called by the state machine at the appropriate time.
You can optionally override the isValidNextState function to control the edges of your state graph, those valid transitions I was talking about.
Now by default, all of these edges are valid but undoubtedly you'll want to use the dynamic internals of your states to decide which of those transitions are valid.
So these four ghost states we talked about: chase, defeated, flee, respawn would all be implemented as GKStates.
So I want to end on an example here.
Let's implement that GhostStateMachine we just talked about.
At the top here I'm going to go ahead and make my four states: chase, flee, defeated, and respawn.
Then I'm going to make my state machine passing in those four states, those are the four states that the state machine can ever be in.
Then I'm going to go ahead and enter the initial state which is chase in this example.
We're good to go.
This state machine is ready for use in our game, and that ghost is going to do exactly what we expect.
So that's state machines, let's move on to agents, goals, and behaviors.
So some concepts before we get started.
What we call agents, goals, and behaviors, are really autonomously moving entities, they're controlled by realistic behaviors and goals.
They have a number of physical constraints, things like masks, acceleration, and inertia.
The behaviors that control these agents are in turn made up of a number of goals, that you combine with the appropriate weights, to achieve some meaningful autonomous movement functionally in your game.
So why are agents so important in game development?
I think a lot of games benefit from really believable realistic movement.
When our game entities move in straight lines and take turns instantly and bump into obstacles in our environment, it doesn't look very real.
Movement in the real world has things like inertia, and mass, and acceleration.
And it correctly avoids nearby obstacles as well as other entities.
And when entities know how to get from point A to B they usually follow a path, and they usually do so smoothly rather than rigidly.
So here's an overview of what we're giving you in our agent system.
We have our Agent class, it is controlled by a behavior and it also has a delegate that let's you respond to changes in the agent.
These behaviors are in turn made up of a number of goals that you combine with weights to achieve that meaningful functionality.
You have a lot of goals at your disposal: things like seeking, and intercepting, avoiding obstacles, and following paths.
Let's go ahead and take a look at the classes.
GKAgent is a simple autonomous point mass and it's also a GKComponent so it plays really nice with our entity and component systems.
And when you call update on GKAgent it's going to apply its current behavior and what that does under the hood it's going to look at the goals that make up its behavior, and calculate along with the weights some total change in acceleration necessary to simultaneously meet those goals as best it can.
It then uses that change in acceleration to change the agents velocity and position in rotation.
Now GKAgent has a had lot of those physical constraints I talked about, things like mass, and A bounding radius, max speed, max acceleration.
It is important to note that these units are dimensionless and very likely to be game world specific.
So you can imagine a game on the scale of kilometers is going to have vastly different numbers here than a game that's on the scale of feet.
So make sure you choose the appropriate numbers for your game world here.
Here we have our GKBehavior class.
And it's a simple dictionary-like container of those goals.
It let's you dynamically modify the behavior as your game undoubtedly changes, and you do this by adding new behaviors, adding new goals, removing existing goals, and modifying the weights on existing goals.
As I mentioned before, you set a behavior on an agent and that agent is good to go.
The next time you update that agent, it's going to correctly attempt to follow that behavior.
So some examples of what behaviors might be, perhaps you want to implement a flocking behavior, to simulate the flocking of birds in real life.
We may combine a cohere, a separate, and an align goal with the appropriate weights to achieve that.
Or maybe I'm making a racing game and want to make a racing behavior to control my race cars.
This might be as simple as combining a follow path, I want my race car to follow the race track, and an avoid other agents goal, I want my race car to avoid colliding with the other race cars.
Here's a code example much what making these behaviors looks like.
You see the top, I'm going to make a seek behavior, I want to seek some enemy agent in my environment.
I'm going to make an avoid goal, I want to avoid nearby obstacles.
Then I'm going to make a targetSpeed goal.
I want my agent to accelerate to and reach some target speed.
Then I'm going to make my behavior passing in those three goals with an appropriate set of weights.
You see here I'm weighting the avoid goal at a 5 because I definitely don't want my agent to collide with nearby obstacles.
Then I'm going to make my agent, initialize it, set the behavior on it, and this agent is good to go.
The next time I call update on this agent it's going to correctly do what I expect.
So let's talk a little about that agent delegate.
GKAgentDelegate is useful when you need to sync your visuals, things like graphics, animation, physics, with this underlying agent simulation.
We give you two callbacks to do that.
agentWillUpdate, which is called before any updates are applied to the agent.
And agentDidUpdate, which is called after the updates are applied.
In your game this might be things like a SpriteKit node, or a SceneKit node, or a render component.
Let's take a look at what this delegation looks like in a SpriteKit game.
You see here I have a custom sprite node MyAgentSpriteNode and I'm going to go ahead and implement both of those callbacks that I talked about.
In agentWillUpdate, I'll set the agent's position in rotation, equal to my node's position in rotation, I want that underlying agent simulation to match my visuals.
Then we are going to do some updating.
And then an agentDidUpdate, I'm going to do the inverse, I'm going to set my node's position in rotation, equal to my agent's position in rotation, the visuals will match that underlying agent simulation again.
Now I would like to give you a quick demo on what agent movement looks look and some of the goals you have at your disposal.
So here I have a simple SpriteKit scene and we're going to represent the agents with a triangle in a circle.
They're oriented where the triangle is pointing.
Here I have a seat goal.
The agent in the center is simply going to try to seek my mouse position.
Notice how fluid and realistic the movement looks, because he's under those realistic physical constraints, things like mass, and acceleration, and inertia.
Here we have an example of the inverse, a flee goal.
The agent is going to instead attempt to run away from the mouse position.
Here is an example of a wander behavior.
My agent is just going to randomly wander about the environment making random left and right turns.
Here we have an example of an obstacle avoidance goal.
Once again my agent is going to attempt to seek the mouse position but I have added a couple circle obstacles to my scene, and he has an obstacle avoidance goal on him.
So while he's still trying to seek the mouse position, he's also going to avoid colliding with the obstacles.
Here I have an example of a separation goal.
I have three agents that are once again going to try to seek the mouse position.
But they also have a separation goal on them.
They're going to try to maintain some minimum separation between them.
And this is really useful for things like formation flying or keeping groups of units together in your game.
Here I have an example of an alignment goal.
The agent on the right is simply going to try to match the heading of the agent on the left.
This is really useful for things like synchronizing units in your game.
Here I have an example of a flocking goal.
Here we have our leader agent in the red which is just going to wander about the scene.
But I also have a group of these blue agents under a flocking behavior.
They are combining a cohere, a separate, and an align goal to stay in a blob, while also trying to chase that leader entity.
So the separate goal is maintaining some minimum separation between them, the cohere goal makes them stay together in a cohesive mass, and the alignment goal wants them to reach an average heading.
Last thing I have an example of a follow path behavior here.
I have a simple polyline path and my agent is going to attempt to follow it.
Now I want you to notice that he doesn't take the corners sharply.
He's under those realistic physical constraints we talked about, things like mass and acceleration.
So he's forced to follow it in a smooth manner even though the underlying path itself is rigid.
So that's agents, goals, and behaviors.
[ Applause ]
Let's go ahead and move on to pathfinding.
Now I'm sure we're familiar with this problem in game development.
I have some entity in my game world that wants to get from point A to B, but there is an obstacle in my way.
I don't want the entity to move through the obstacle.
I don't want her to bump into the obstacle.
I want her to correctly find a path around the obstacle like a human would.
What I'm looking for here is something like this: I want her to correctly find the shortest way around the obstacle, clear the obstacle, and continue on to my goal.
This is the realm of problems we call in gameplay pathfinding.
Now some concepts before we get started, pathfinding operates on a navigation graph.
In this navigation graph, it is a collection of nodes that describe the passable areas in your game world.
The places where my entities are allowed to be and move.
These nodes are in turn joined by connections to describe how my entities move between these passable areas.
And these connections can be single directional or bidirectional, and there is always exists an optimal path between any two nodes in a connected graph.
And this is usually the path we're looking for in pathfinding.
So let's go ahead and take a look at the classes.
Here we have GKGraph, which is our abstract graph base class, it's quite simply a container of graph nodes, those descriptions of the passable areas in my game world.
It has the functions necessary to add and remove nodes as the game world undoubtedly changes, and it also lets me connect new nodes to the graph, making the appropriate connections to existing nodes I would expect.
Add of course we also let you find paths between nodes and a graph.
And we're offering you guy's two specializations, a GKGraph that works with grids, and a GKGraph that works with obstacles.
Let's talk a little bit more about those now.
This is our GK graph that's specialized for a 2D grid.
And what this does, is it's going to automatically create all the nodes to represent a grid of some given start position and width and height.
It's going to automatically make the cardinal connections between the grid nodes and optionally the diagonal ones as well.
And it also has easy functions available to add and remove grid spaces as they undoubtedly become impassable and passable again in your game.
Next we have our GKObstacleGraph.
This is a GK graph that's specialized for pathfinding around obstacles in your game world.
And these obstacles can be any arbitrary polygon.
Now we give you the functions necessary to dynamically add and remove obstacles as your game world, again, undoubtedly changes.
It also lets you dynamically connect new nodes to the graph and this is really useful for stuff like inserting a start and an end node in my graph to find a path for a unit.
Now we do this by what we're calling a buffer radius, this is a safety zone around obstacles, where my entities are not allowed to go, and it's often a game-dependent size relating to the bounding radius of the entities that I want to do the navigating.
So let's talk a little more about how these obstacle graphs are generated.
So here I have a simple scene with two square obstacles, an entity on the lower left that wants to get to that bridge on the lower right.
My entity is bounded by some bounding radius, and we're going to use that as our buffer radius to artificially make our obstacles larger.
Then under the hood the obstacle graph is going to make the appropriate connections between all of our grid nodes, and it's going to correctly not make the ones that would violate the spatiality of our obstacles.
So you see here that we found that shortest path we were looking for.
It doesn't collide with any of my obstacles.
Here is a code example of that last example, but with just a single obstacle.
Here at the top I'm going to make a simple square polygon obstacle; it's just four points.
Then I'm going to make our obstacle graph, passing in our obstacle and some buffer radius.
Then, I'm going to make a start and end node.
One for where my hero currently is and one for where she wants to go.
Then I'm going to dynamically connect those nodes to my obstacle graph using the obstacles that it possesses.
And what it's going to do is it's going to insert those nodes into the graph and again automatically make the connections that make sense, and not make the ones that would violate the spatiality of my obstacles.
Then at the end here I'm going to find a path for my start and end node and I get back a simple NSArray of graph nodes, which I can then use to animate my character.
Some advance nodes on our graph node class which is GKGraphNode, undoubtedly some need will arise where you want to subclass this.
And this is really useful for implementing stuff like advanced or non-spatial costs or for when you need more clear control over the pathfinding.
You can imagine a strategy game that has a variety of terrain types.
Perhaps you want a forest terrain type to take double the move over as my other terrain types.
I correctly want pathfinding to take this into account.
I don't want it to return the visually shortest path.
I correctly want it to navigate around the forest.
Because that is actually the shortest path in my game world's terms.
GKGraphNode is also useful when you want to manually make your own graphs, and you do this by manually managing the connections between nodes.
This is really good for things like abstract or non-spatial graphs.
Perhaps you want your game to have portals and your units to correctly take those portals into account for pathfinding purposes, even though those portals aren't spatially connected in anyway.
And our Grid/GraphNode and GraphNode2D which is used with our obstacle node are also available for subclass as you see fit.
This is a feature I'm really excited about.
We have done some work with the SpriteKit team to allow you to easily generate these obstacle graphs from your existing SpriteKit Scenes.
And you can do this for things like node bounds, node physics bodies, and nodes textures.
So what this means is with very few lines of code you can take an existing SpriteKit scene and generate an obstacle graph and automatically pathfind around it.
Now I would like to give you a small demo of this.
Let's explore pathfinding with SpriteKit.
Here I have the tower defense game we have talked about implemented as a SpriteKit scene.
I'm generating entities on the left and they want to move to the bridge on the right.
But because this is a tower defense game I'm undoubtedly going to place some towers right, that violate their current path.
So let's go ahead and place one.
And you'll notice they correctly pathfind around it.
That's because we're using the SpriteKit integration, that we just talked about, to automatically generate an obstacle from that node, update the underlying GKObstacleGraph, and update our path.
So let me turn a debugger, let me remove this tower real quick.
You see we just start with our simple path, right, between start and end node.
But as I insert an obstacle in here, we recalculate the underlining GKObstacleGraph.
And this allows our entities to find a new path around that obstacle.
So let's go ahead and add a few more.
And because of that SpriteKit integration, every time we add or remove an obstacle, we can keep that underlying GKObstacleGraph updated.
So that's pathfinding with SpriteKit.
[ Applause ]
Now I would like to call my colleague Ross Dexter up to tell you a little about our MinMax AI.
[ Applause ]
ROSS DEXTER: Thanks, Bruno.
So many of the features that Bruno spoke about earlier can be used to create AI, but they're more about giving life to entities within your game.
Many games also need equal AI opponents that can play the entire game by the same rules as human players.
And this is critical for games like Chess, Checkers, Tic-Tac-Toe, and so on.
So we wanted to provide you a solution for this.
And so we've chosen to implement a classic AI solution, MinMax, as a key part of GameplayKit.
MinMax works by looking at all the moves available to a player, and then it builds out a decision tree, from each of those moves and all the permutations that can arise from each of those moves.
When you request a move from it, it searches this, the decision tree, looking for a move that maximizes potential gain while minimizing potential loss.
This Tic-Tac-Toe example here, the AI selects the move on the right for the X player because in the best case it results in a win, or in the worst case, it only results in a draw, the other two moves both lead to losses.
So MinMax AI gives you the ability to add AI controlled opponents to your games, but it can also be used to suggest a move for human players when they get stuck, and it's going to be great for games that even don't have any other AI requirements.
It's best suited for turn based games, but it can be made to work with any game where you have a set of discrete moves available for players to make.
You can easily adjust the difficulty of the AI by varying how far out into the future it looks.
Looking 10 moves in advance results in much more effective play than looking ahead only 2 or 3 moves.
Additionally you can optionally direct it to randomly select suboptimum moves to give it an element of human error.
So let's look at how this integrates with your game.
The great thing about MinMax is that it doesn't need to know any of the details of your game.
You don't need to teach it your rules and it doesn't need to know how it's implemented.
This is all abstracted away.
All you have to do is provide a list of players in the game, the possible moves they can make, and a score for each player that indicates the relative strength of their current position.
When you request a move from the AI, it takes all this data into account and it builds a decision tree, and returns the optimal move for you to use.
Let's look at the classes.
There are three key protocols that you're going to need to implement to work with the MinMax AI.
And the first of these is GKGameModel, and it's an abstract of the current game state.
If you're creating a Chess game for example, a good candidate to implement this class would be on, say, the board class because it tracks all of the positions on the board as well as all the pieces that are currently in play.
As I mentioned on the previous slide, all this needs to do is provide a list of the players that are active in the game, the current player, scores for each of those players, and then the possible moves that each of those players can make.
It also needs to have a method for applying those moves and this is used by the AI to build out its decision tree, and can be used by you to apply a move after it's been selected by the AI.
And when this move is applied, it will change the current game state, possibly changing the current act of the player, scores for each of those players, and the moves that are available to them.
The next protocol is GKGameModelUpdate, this is an abstract of a move within your game.
It should have all of the data you need to apply a move to your game model.
As we have said, it is used by MinMax to build out the decision tree, and can be used by you to apply a move after it's been selected.
Finally we have GKGameModelPlayer which is an abstract of a player of the game, and it's used by the AI to differentiate moves from one another.
Now we get to the AI itself, it's within the class GKMinMaxStrategist, and it operates on a GKGameModel.
So after you create an instance the MinMaxStrategist, you're going to hook it up on the gameModel property.
maxLookAheadDepth is how far into the future it looks when you request a move from the AI.
And as we mentioned earlier higher numbers result in more effective play than lower numbers.
And that's all you need to do to start using it.
When you call bestMoveForPlayer, the AI will build out its decision tree, rank all the available moves in order from best to worse, and then return the optimal move.
There may arise cases where you'll have more than one move that is equally advantageous for the AI to make, and in those cases you can direct the AI to randomly break ties.
And that sort of thing comes in use if you want to call randomMoveForPlayer.
Say you have 10 moves available for a player, but you only want to select a random one from the 3 best moves, it will take that sorting and randomly choose one of those 3 best moves.
One of those moves may be suboptimal unfortunately, but that may be desirable if you are trying to make your AI appear more human and have a chance of making an error.
And both bestMoveForPlayer and randomMoveForPlayer return a GKGameModelUpdate which you can then use to apply to your GKGameModel to make a move.
So here is a quick code sample.
Here we're creating a Chess game model.
And unfortunately going over the details of how you might want to implement your game model are beyond the scope of this session, but we do have excellent sample code available that you can look at to show how you might want to go about doing this.
So we create our Chess model, and then we create our MinMax AI, and hook it up by setting the game model on the gameModel property.
We then choose our LookAheadDepth to 6, so we're going to look ahead 6 turns in advance when we build our decision tree.
That's all we need to do.
Now we call bestMoveForPlayer with a currently active player and it will find the optimal move for that player with the given information.
You can then apply that move to the game model to make the move.
So let's look at a quick demo.
So here we have a simple game where there are two players, black and white, and they're trying to get as many pieces of their color on the board as they can.
When they place a piece on the board they will flip any colors of the opponent's pieces to their color that lie between their own pieces.
So here we have both players controlled by the AI, the black player is looking ahead five moves in advance, while the white player is only looking ahead three moves.
This allows the black player to easily defeat the white player as it goes through the game.
You can see here we have a score for each player.
This is simply we take a look at how many pieces the player has on the board minus the number of pieces that their opponent has on the board, adjusted with some weights, and that gives us our score.
So you see here the black player easily defeats the white player.
So let's look closer at the score here.
We see here that we have all of the pieces in the center, they're weighted at 1.
The position on the edge of the board are weighted higher, the corners are weighted even higher.
That's because these positions are more advantageous for the players, and so we direct the AI to favor these places by changing how those places effect the scores.
So let's change-up the look ahead on these guys.
We'll make white look ahead 4 instead of just 3.
And even just this small change will allow the AI to play more effectively and in fact in the middle of the game it looks like the white AI has the upper hand, but the black AI is able to trade a short-term game for a long-term victory, and is able to overcome white in the end.
That's MinMax AI.
[ Applause ]
ROSS DEXTER: So now let's talk about random sources.
And at first this topic may seem unnecessary, because we already have rand.
Why shouldn't we just use that?
Well rand gives us random numbers but games have unique random number needs, and rand may not give us everything we want.
First of all the numbers that rand generates may not be the same from system to system.
You're not guaranteed to have the same results on different platforms.
And that can be a big problem for networking games, because if we can't rely on the numbers on either side of the collection to be generated in the same sequence we have to waste critical bandwidth in syncing those two sides up.
So we want platform-independent determinism.
Also whenever we make a call to rand we're drawing from a single source.
So if I have a bunch of calls to rand in my AI code, and then I add a new call in my physics code, that call in the physics code will affect the numbers that are being generated in my AI code, which could result in unexpected behavior.
What we really want to do is be able to separate those two systems apart, so that the numbers generated in one system have no effect on the numbers generated in a different system.
And also we may not want control over just the range of numbers we're generating but also how those numbers are distributed across that range.
And this is where random sources comes in.
So we're offering you a set of game quality random sources that are deterministic, so when you have the same seed, you will always get the same sequence of numbers no matter what platform you're on.
They are also serializable so they can be saved out with your game data.
And this can be really useful in helping to prevent cheating.
And they're also implemented using industry-standard algorithms that are known to be reliable, and have excellent random characteristics.
In addition we offer you a set of random distributions to leverage, and these allow you to control how your numbers are distributed across the given range.
We have a true random where every value is equally likely to occur, Gaussian distribution where values are weighted on a bell curve with values toward the mean more likely than those on the fringes, and also anti-clustering or fair random distribution which helps eliminate runs of numbers.
And finally we have NSArray shuffling, which is super useful for doing things like shuffling a deck of cards.
So let's look at the classes.
GKRandomSource is the base class for random sources.
And it adopts NSSecureCoding and NSCopying so it can be securely serialized.
Determinism is guaranteed with the same seed, no matter what platform you're on, so if you want the same sequence of numbers, you can always rely on it to be generated.
If no seed is given, one is drawn from a secure system source.
Go on to sharedRandom, which is the system's underlying shared random source, and this is not deterministic but there are cases in which this may be desirable, such as when you're shuffling a deck of cards and you want every result to be unique.
Let's go over the AI random source algorithms we have available for you.
We have ARC4, which has very low overhead and excellent random characteristics and is going to be your Goldilocks random source, we have Linear Congruential which has even lower overhead than ARC4, but it's random characteristics are not quite as good, and you may see some more frequently repeated sequences of numbers, finally we have the Mersenne Twister, which is high quality but memory intensive.
Note that none of these are suitable for cryptography but Apple offers other separate APIs to meet these needs.
Now we get to our random distributions, in the base class for this, it is GKRandomDistribution which is implements a pure random distribution, meaning every value between lowest value and highest value, are equally likely to occur.
You can get numbers by calling nextInt, nextUniform, and nextBool.
We also offer a set of dice convenience constructors to create 6 sided, 20 sided, and custom sided die.
Then we have GKGaussianDistribution which implements a bell curve Gaussian distribution.
The values are biased towards the mean value and the values farther away from the mean are less likely to occur, and that's what happened in our sample distribution here.
We have generated a sequence of 15 numbers between 1 and 5, and we see that the mean value of 3 occurs far more frequently than any of the other numbers.
In fact it occurs more than twice as frequently as any other number.
With 1 and 5, the values on the fringes, only occurring a single time each.
Note that in a standard Gaussian distribution, is unbounded but that's undesirable for a random source, so we call every value outside of a three standard deviations of the mean.
Next we have our anti-clustering distribution implemented in the class GKShuffledDistribution.
This is our fair random distribution, which helps reduce or eliminate runs of numbers, but it's random over time.
And you control this by using the uniformDistance.
At 0.0, all numbers are equally likely to occur, and this is indistinguishable from a true random source, our random distribution.
At 1.0, all values are different and it will run through every value in the range before you start to see any repeated values.
That's what we have here.
In our distribution here.
Once again we're generating 15 numbers between 1 and 5 and you can see that we're hitting every number in the range before we start to see any repeated values.
And in fact every value is generated exactly three times.
So let's go over the simple code examples.
It's very easy to create a 6 sided die random source, you just use the connivence constructor GKRandomDistribution, and rolling the dye is as easy as calling nextInt.
It's similarly easy to create a 20 sided die.
And creating custom die is also quite easy.
Here we're creating a 256 sided die which would be pretty awkward if you tried to roll it in the real world.
The previous three examples were all implemented using a true random distribution, but you can use any of the distributions that we have available to you.
Here we're creating a 20 sided die with a Gaussian distribution, so it's weighted to the mean value, around 11, so when you roll it, you're most likely to come up with a number around there.
And here we're creating a die, a 20 sided die with our shuffle distribution, and by default the uniform distance on our shuffle distribution is 1.0.
So when we roll this one, we're going to hit every value in the range before we start to see any repeated values.
The first time we roll it, we might get 5, then we know the next time we roll it, we definitely not going to get that number again, until we run through every other value in the range.
And finally, here we have array shuffling, we're using the shared random source we mentioned earlier on GKRandomSource, which gives us access to the system's underlying random source, which is not deterministic, but in this case that's advantageous.
We want every instance of the card shoveling to be unique.
And you can see how easy it is to make random sources a part of your game.
It's only a couple lines of code and you can get going.
And that's random sources.
So now I would like to invite Joshua Boggs up here to talk about our rule systems.
[ Applause ]
JOSHUA BOGGS: Thanks, Ross.
Hi. I'm Josh.
I have been working alongside Bruno and Ross while they've been putting on the finishing touches to GameplayKit.
I'm here to talk about one of those systems, the rule systems.
So before I go into the rule systems, I just want to go over some common ingredients that games tend to have.
Games tend to consistent of three elements, it is things like your nouns: position, speed, player health, equipment they may be holding.
Secondly, you've got things like verbs: these are actions that the player can perform, things like run, jump, using an item, or if you're in a car, accelerating.
Lastly, the rules.
Rules are incredibly important because they define how your nouns and verbs interact.
Rules give flavor and texture to your gam, and great games have great rules.
So let's have a look at an example rule.
Here we have a rule that a driver may use to decide when to brake and when to accelerate.
Using an input property of distance the player will either slow down or speed up.
We can see in this example that if the distance is less than 5, they're going to brake, when it's greater than or equal they'll accelerate.
This is fine logic, but there is a subtle problem.
In the distances around 5, we're going to get very jerky movement because the car is going to continue to oscillate between braking and accelerating.
This is going to give us very jerky movement.
So for more natural movement we need something a little more approximate.
Using a more fuzzy solution we output facts about what to do rather than perform the actions immediately, we've output two facts here, closeness and farness, both based on distance.
The important thing is you can now be both close and far.
So rather than perform one or the other, this lets us blend the two together to get a more natural movement.
This is especially important around the previous example.
Now when the distance is around 5 we'll get much more natural acceleration.
This is the motivation behind rule systems.
Facts can be grades of true.
This allows us to perform more complex reasoning with fuzzy logic.
Fuzzy logic deals with approximations.
It also allows us to separate what we do from how we should do it, rather than performing actions immediately, we just state facts about the world, and then take deferred actions later based off of those facts.
So let's take a look at one of those classes.
Here we have GKRule.
GKRule consists of a Boolean predicate and an action.
The predicate matches against facts and the state in the system and only fires its action if the predicate is true.
Actions could be as simple as asserting a fact, or as complicated as you'd like with a complex block.
Importantly they can now be serializable using NSPredicate serialization methods.
The important thing to remember is that rule systems provide approximations to answers.
Things like how close am I to the car in front?
In the first example we can kind of say that with a fairly high grade of confidence, that we're quite far.
Where with the other two, things are a little more fuzzy, answers that we're after, things like somewhere in between, closer.
Let's have a look at the system that manages these rules.
Here we have the other class, GKRuleSystem.
GKRuleSystem is an ordered collection of rules and facts.
To assert facts about the world, simply call evaluate on it.
This will to run through the rules in the array and those rules will use a state dictionary as input and insert facts later based off of that.
The facts will be held in the facts array and it's important to know that whenever a fact is asserted the evaluate will actually go back to the beginning, and continue evaluating.
This is because when you assert a fact, this may affect the way other rules work.
This ensures that when evaluate is finished you know you have the most concise and accurate view of the game.
To start over again, like maybe at the end of an update loop or on a timer, simply call reset and will clear up old facts so that you can repeat the evaluation.
Let's have a look at the code example.
Here in the beginning, we initialize our rule system, and then later we access the state and assert two facts based off this.
Later in the game code, excuse me.
We then grab these two grades and sum them together to get a sort of fuzzy approximation about how much we should accelerate, and feed this in our game code.
So let's take a look at little example we have going.
Here we've got cars driving along the freeway.
The cars in the intersections are using one set of rules, and the cars on the freeway are using a different set.
The ones on the freeway are deciding how much they should slow down or speed up based off the distance of the car in front.
They're asserting two facts about the world.
These are things like distance, relative speed.
The cars in the intersection are using a different set of rules and asserting facts on who has the right of way.
Putting them altogether we can get very complex simulations about the world.
This is a power of rule systems.
So before I go just some best practices on using the rule systems.
It is important to remember that GKRuleSystem is isolated.
You should be using the state dictionary as a snapshot of the game world.
You should also use many simple rules and assert many facts about the game world as opposed to large complex rules and fewer facts.
It is also important to note that facts are approximations and it is up to you to decide how you should use them.
The grade of a fact is the system's confidence in it, and this allows us to use fuzzy logic to achieve more complex reasoning.
With that, I would like to hand it back to my colleague Bruno to finish up.
[ Applause ]
BRUNO SOMMER: Thanks, Josh.
So that's GameplayKit.
Today we talked about the seven major systems in GameplayKit, entities and components which are a really great way to structure your game logic.
State machines which deal with the statefulness in our games and the various state changes that our objects undergo.
Agents, which are our autonomously moving entities controlled by realistic behaviors and goals.
Pathfinding, which deals with navigation graph generation and finding paths within our game world.
We also talked about our great MinMax AI solution, which helps you give life to your computer controlled opponents.
Also the slew of great random sources and distributions that are available to you.
Lastly we talked about rule systems which are a great way to describe discreet and fuzzy logic.
We really are excited to finally get GameplayKit in your hands and can't wait to see what you make with it.
Some great code samples dropped this week, you should definitely check it out if you want to learn a little more.
DemoBots is a SpriteKit game covers a wide variety of the GameplayKit API, FourInARow is a good example of MinMax AI in action, and AgentsCatalog is a really good example of the agent's behaviors and goals, so definitely check that out if you want to learn a little more.
There is also some sessions coming up if you want to find out a little more about our related technologies, SpriteKit, ReplayKit, Game Center, SceneKit.
After lunch today we have a deeper dive into DemoBots which is that sample I talked about, so definitely check that out if you want to learn a little bit more about GameplayKit or SpriteKit.
There is also some great labs coming up, check out the Game Controllers lab.
There is also a GameplayKit lab today after lunch, meet the team, ask questions, talk about any problems you might have with the code.
If you need anymore information, we direct you to check out our great developer site and for any general inquiries contact Allan Schaffer, our Game Technologies Evangelist.
Have a really great rest of your conference.
[ Applause ]