Skip navigation

Tag Archives: game ai

Since I spent some time tweaking the AI in Freekick 3, I thought I should write down some of the implementation details for future reference.

The AI has quite a simple general structure. Apart from the special cases like restarts, goalkeeper AI etc. the main gameplay AI consists of three states: defensive AI, offensive AI and the AI when the agent is about to kick the ball. Whether a player is in the defensive or offensive state depends on the match situation (where the ball is, who has the ball) and the team tactics (agent’s position).

The defensive state is basically a decision “where should I stand to make the opponent’s offensive difficult”. There are a few options: try to take the ball from the opponent, try to block an opponent’s pass or shot or guard some area or opponent. The AI simply assigns scores to each action and picks the action with the highest score. (This, I suppose pretty standard AI technique, seems to be inspired by utility functions and is used throughout the Freekick 3 AI.)

The offensive state is even simpler than the defensive state: either the player tries to fetch the ball or tries to place himself in the best possible supporting position, which should be somewhere that can be passed to, far away from opposing players, and a good shot position. (The AI builds a kind of an influence map that is also affected by some soccer-specific things such as the offside rule.)

Probably the most important decision the AI has to make is what to do when the agent has the ball. Again, like with the defensive state, the AI has a few different possible actions, and it assigns scores to all of them and simply picks the action with the highest score.

The possible actions are Pass, Shoot, Long Pass, Clear, Tackle and Dribble. I’ll start with the Pass action.

When deciding the pass target, the AI loops through all of the friendly players and keeps track of the best option. For each player that’s not too far or too close, the AI considers either passing directly to the player or trying a through-pass. The base score for the pass is highest for players nearer the enemy goal, and then decreased for each opponent player that is seen as a possible interceptor.

The Shoot action score is basically a function of distance to the opponent’s goal and the distance from the opponent’s players (especially the goal keeper) to the possible shot trajectory.

The Long Pass action is actually a composite of Shoot and Pass – it checks for the Shoot and Pass action scores of the friendly players and chooses to make a long pass (or cross) to the player with the highest score. As with other actions, the score is multiplied by a team tactics coefficient, allowing the coach to influence the team’s playing style.

Clear and Tackle are basically emergency brakes that the AI can pull in a situation where the ball needs to be kicked away from the own goal or the opposing players.

With Dribble, the AI creates a few possible rays at regular angle intervals around the agent as possible dribble vectors and assigns scores to each of them. Similar to shooting, the score is higher near the opponent goal, but decreased by opponent presence.

So, in the end, the AI is composed by several rather simple techniques. It’s all heuristics without any algorithms providing optimal solutions (if any can even be used in soccer AI). It uses some simple seeking behaviors (arrive at a position, chase the ball). There’s one simple state machine, with state transitions decided mostly by the ball position. The top level AI is built around simple if-then-else-statements (I suppose you could call them decision trees). Lots of the decision making uses some sort of fuzzy logic, even though it’s not really structured like that (instead the code itself is fuzzy). Still, the AI manages to seem smart in most cases, it plays rather well and presents a challenge for the human player (for me, at least).

There are still quite a few standard AI techniques that I haven’t implemented which might make sense for Freekick 3. For example it might be interesting to experiment with adding some kind of learning ability to the AI, which should be possible using reinforcement learning, or adding a more complicated planning process with the use of a decision network. A useful first step would be to extract all the used generic concepts like decision trees and fuzzy logic to their own code pieces, which would enable experimenting with things like learning a decision tree.

My conclusion is that there are lots of different game AI techniques, many of which are quite simple, and the key to creating a fun AI is to find out which techniques to use for which problem and how to combine the techniques. The techniques are often intuitive but can be also described mathematically, so that when reading up on game AI, you may, depending on the material, get an impression that game AI is either very non-scientific or very mathematical (and therefore difficult), while I think it can be either, depending on how you look at it.

For the interested, the ~500 lines of Freekick 3 AI that make up the core can be found at GitHub.

About a month ago I announced my then-latest programming project, Star Rover 2. It’s a space flying game written in Haskell, and while programming that was fun and I learnt a lot from the project, it was time to find something else to do. So I started another soccer project, again in Haskell (like in September 2008), but with a bit more Haskell experience under my belt.

And now, after almost a month, it’s time for the first release. The new soccer game is called freekick2, and it’s an open source 2D arcade style soccer game.

If you have cabal, you can install it with a simple cabal update && cabal install freekick2. Here’s a screen  shot from a menu:

The freekick2 distribution includes only a few fantasy teams, but you can use the data files from a Sensible Soccer game for some additional realism.

One lesson I learnt when doing Star Rover 2 is that I should avoid over-engineering like plague, as that was probably one reason why I couldn’t get the first Freekick versions in a releaseable state.

Freekick2 now consists of about 3200 lines of Haskell code, around 1000 more than Star Rover 2, and it’s also a bit fancier, graphics-wise. With Star Rover 2 I used monad transformers to handle the main loop and state changes, which worked pretty well, and I’m using the same technique with freekick2. I still haven’t figured out the whole FRP thing, but game programming in Haskell seems to work out pretty well without.

Things I’ve learnt while programming freekick2:

* Texture mapping with OpenGL in Haskell. To load the textures you need to play around with pointers a bit, which are not really that usual in Haskell. But you can do it, and with the help of some great small libraries like pngload it’s relatively simple in the end. Adding dynamic recoloring of the textures after loading them (to support multiple kit colors) was fun as well.

* Parsing binary data in Haskell. Over a year ago, when I was writing Freekick in C++, I wanted to deserialize the SWOS data files (the format’s known; many thanks to the ones who figured the format out) and I thought parsing binary data sounds like a job for C or C++. The result from back then is about 850 lines long and took a few hours to write. Granted, it also generates some XML. Now I wrote a program to do the same task in Haskell; the result has 250 lines and was written in less than two hours. The Haskell version also includes code for writing SWOS data files – a feature I added just because it was so easy to do. Later, for the freekick-internal data structures, I let BinaryDerive derive the serialization/deserialization code for me. Needless to say, I won’t be doing binary data parsing in C anymore if I get to choose.

* Template Haskell. The state in freekick2 changes pretty often, and is managed in a State monad. I needed a lot of functions of type (FieldType -> FieldType) -> State -> State (like already with Star Rover 2), usually one for each field in a record, and so I let TH generate them all for me. Bulat’s tutorials (links to them can be found here) were a life saver.

* Implementing a GUI with colors, textures and dynamic content, with a tree-like hierarchy for screens.

* Use of the existential quantification language extension for heterogeneous collections. I’ve spent some time with object oriented design in the past, which sometimes leads to object-oriented designs in Haskell, so that something like existential types are needed. In this concrete case, I created the class Sprite, with both the player and the ball implementing it. The easiest way to draw all sprites sorted by depth is then to just put all sprites in one list and sort it. Without existential quantification, I would’ve had to write code to sort the players and then call the draw method of the ball somewhere between the draw methods of the players.

So, programming in Haskell is more fun than ever, and the productivity boost given by Haskell is starting to show more and more.

Time for a small update again. The last time I wrote I mostly reviewed the work done last year, wrote about the tools used and spent some time comparing Haskell with C++ as I had been busy rewriting the Freekick server in C++. So what’s been going on in the past month? Well, some of the points mentioned in “things to do next” part of the last post have been worked on, while others aren’t. Also, as I have now been using C++ for a few months there have been some signs of the need to wander off to another – hopefully better and shinier – programming language again, just like one would expect. This time I’m not thinking of Haskell though, but Python.

I’ll first let you know what’s new with Freekick core development. There’s been a lot of small improvements – I’ve added some server functionality such as allowing human players to play along or goal detection, but the most important update is that the OGRE client now has the ability to control a player and kick the ball around. This actually makes Freekick a soccer game that can be played. This doesn’t mean, however, that it would actually be fun to play the game – the relatively silly AI makes sure we’re not quite there yet. Another thing that makes the game less fun is that the clubs aren’t really there; the players have no differences in skills or personality, there are no kits or club names or tournaments.

The next thing for me to do is try and eliminate these two problems: first, the AI needs to be made better. At the moment it’s not really that bad – they pass the ball and try to score – but they only have a few actions which makes the match seem repetitive, and the goalkeepers are horrible. Second, the soccer match needs to have a menu around it, including creating and viewing tournaments, lineups, league tables – the whole deal. Actually I’ve already started work on that, which brings me to Python.

You see, working with C++ the whole time with Freekick core was not bad, but after you’ve seen something like Haskell in action, sooner or later you start to wonder if everything really needs to be so complicated. I’m talking about the sheer amount of time spent typing the code in, the loops, the iterators, the classes and headers and declarations, the curly brackets…

As Freekick was designed with modularity in mind and was split to multiple processes from the beginning on, the question of using Python for the game menu arose. After poking around with PyQt for a while the first drafts of the game menu were already finished. I had had my doubts about dynamic languages and duck typing before, but now, after having written a few small applications with Python I’m slowly starting to see the advantages such a dynamic high-level language can bring. I was amazed how fast you could actually write a simple-looking Python script that does the same thing as an application you could spend weeks writing in C++.

Encouraged by my Python adventures, I played around with the idea of solving the other problem – AI – with Python. Since I don’t really fancy writing all the Freekick client side logic in Python, I thought about using Boost.Python to export the C++ classes into Python. Doing the experiment was interesting, but the unfortunate truth is that exporting the classes still seems like too much work, even with the excellent library, especially with the page-long error messages and the fact that the STL containers would apparently also be have to exported manually. The other way would be embedding Python in C++, but I’d still have to export the data structures to Python and back, which makes me think it’s probably faster just to stick with C++ after all.

The GUI part is still far from ready – the only event that starts a match at the moment is a friendly match, and the GUI-match interface is non-existent at the moment – but the base is there and will be expanded when the time is ready. At the moment I think improving the soccer match itself is more important, since there is really no good reason to start a match if playing it isn’t fun, which brings me back to C++ and the AI. I’ve been reading the book “Programming Game AI by Example” by Mat Buckland again, which really is quite an inspiring work, but I think it will still take a while for me before the AI code is in a state where it both seems halfway smart and is easily extensible.

Luckily there are still a couple of mini projects for me to do in case writing AI gets boring. First of all there’s the GUI with all the tournament and schedule creation to do, but I think the time is also slowly getting ready for another Freekick client, this time with ncurses as the “graphics” library. I’ve been wanting to get my hands dirty with ncurses for a while now, and I’m all for minimalistic software, so hey, why not. Other fun things include creating a TV style soccer camera in OGRE and taking player skills into account on the server side. I’ll keep you posted.

In my last post I described the high-level architecture of my software, how the soccer match part is divided into a client that does the work on showing the graphics and receiving input, a server that actually has all the match info and the AI process that actually acts as a client but controls more players at once. I also said my next task would be to write a small working piece of code for each task to get me started. Well, I’m still there; the main physics part is done. As an exercise to working with the BSD sockets interface in another language than C I did the physics part in Haskell, which was fun. Fun, but not easy.

Basically the difficult part was that I wanted the server to establish connections with many clients, send them some information about the match during connection and then broadcast the match data to all clients simultaneously. For this, of course, you need concurrency and threads, though the fact that Haskell as a purely functional language already threads everything confused me a bit. The answer is Control.Concurrent.forkIO and, for the more complicated stuff (i.e. communication between threads) Control.Concurrent.STM. For tutorials on how to program servers with Haskell, a nice, light introduction can be found at while a bit more difficult but certainly as interesting article is at For a client, check out But I digress. Actually I wanted to write about AI.

So, after finishing my physics server (actually it’s not completely finished, it has no soccer rules yet, but it’s a server with some physics understanding after all), I went on to do the AI part. Since it can get quite complicated, I thought about doing some exercises on AI first. Now, the way how to start designing and programming AI was not really clear for me. I have the book “Programming Game AI by Example” by Mat Buckland, (a very nice book by the way) which has installed the relatively simple scheme of state machines in my head. I was not all sure, however, if a state machine would really be enough for soccer players, so I went on to do some more research.

I then stumbled upon neural networks, a topic about which I had only briefly read in the past – they seemed interesting but I never had the patience to sit down and try to grasp the theory. I finally went through the basics about them on a few sites including Wikipedia, but noticed none of those sites really showed any practical examples. I finally came upon two web sites that really explained how to use neural networks in detail. The first one is actually in German and describes the use of neural networks in a Counter-Strike bot, see It was a very interesting read, but for grasping the details you’d need to dive in the source code. The second site can be found at; this one is a very very practical introduction to neural networks, it helped me a lot and it’s also by Mr. Buckland, who keeps surprising me how simple it can be for some to make difficult things seem simple. His tutorial trains the networks with genetic algorithms, for which he also wrote a great tutorial ( The code, like all code from Mat, seems to be written for a less standards compliant compiler such as Visual Studio 6.0 or similar, which means you have to tweak it a bit if you want it to compile on a compiler that follows a C++ standard, but it’s doable.

Encouraged by the theory and these practical tutorials, I went on translating Mat’s Minesweepers into Haskell. During this process I also realized why almost all of the tutorials and the theory of artificial neural networks didn’t show any practical examples; my neural network code itself is merely about 50 lines of active code and contains mostly just the mathematical formulas to the networks. The applicaton of neural networks, however, is nearing 200 lines of code (huge, isn’t it?) and basically describes the whole learning and using process of the brain, err, neural network.

What I learned was that how the neural networks work is basically just a few formulas that don’t really say much about the application; in the end the network is really just a black box that you don’t really know, but it just “magically” gives you the correct values. The difficult part when applying neural networks to a problem is designing the way they’re used and figuring out how to make them learn as well and fast as possible. (I haven’t figured out any of the other ways of getting networks to learn (like simulated annealing and whatnot), i.e. providing positive feedback to them than genetic algorithms, but for now it seems good enough.) I also realized how the whole process was based on randomness, and while I thought that was quite funky, I did have my moments of despair trying to figure out if the thing had actually learnt something or not. (If in doubt, wait.)

Well, in the end it worked, though the minesweepers weren’t rolling around the screen like in the Windows/C++ counterpart, but the path of a minesweeper is drawn into a SVG file instead, as in In this picture, you can actually see a minesweeper (the line) going for the mines nearby (the dots). Oh, want to take a look at the code? That can be found at

That being cleared out, the next thing to do would be to actually finally implement a simple soccer AI. And, well, if I actually did it like I originally planned, with the locations of all the players and the ball and the player skills and personality as input, action including the direction and power of the action (let it be running, kicking the ball etc.) as output, it would probably end up being a pretty big neural network. The time to train it would probably also be pretty long, I guess.

The solution I’m trying to come up with includes state machines, neural networks, decising trees and maybe an occasional Markov decision process. (I also need to finally get my hands on the book “Artificial Intelligence: A Modern Approach”, which I’ve managed to miss… until now.) The basic principle is the same one as discussed in my last post – breaking everything into smaller parts, with state machines or decision trees to assist with the splitting, while using neural networks for the smaller, more limited problems. Probably my next post will still be on AI, but when that’s done (or at least the skeleton of it), and the AI players are kicking the ball, the only bigger thing missing is a client that actually shows us the action…