Using AI for Game Design!

hey this is Jonas making games with you and today we've got some very interesting questions to tackle can we improve an ancient board game using artificial intelligence and how can you write an AI that can play almost any board game at a superhuman level because yes that is very possible and can we use AI as a game design tool to design better games in general some very interesting questions and the game we're going to look at today is this one who bow it's a traditional African board game it's also the favorite game of my girlfriend so I've got to be a bit careful what I say about it and it's definitely an interesting game would have made for a great video by itself probably so let's go have a look at how it plays from a game design perspective let's open the game up this is your side of the board and this is your enemy side of the board and the goal is to get all of the stones to your side when you make a move you grab all stones from one hole then you move clockwise dropping one stone into every hole when there are any stones in the last hole you drop the stone into you grab all the stones in that hole and repeat the process until you end in an empty hole

that's when your turn is over after that your enemy performs a move on his side of the field and one little extra rule is just that you cannot start a move from a hole with only one stone you need at the very least two stones the way you steal your enemy Stones is by Landing in a filled hole of your front row in that case you're allowed to grab your enemy Stones as well and continue with all of them as you can see here this can even happen multiple times per turn and you're also allowed to steal the stones from the back row you're not allowed to steal from your enemy if the front row is empty in that that case the back row is protected when it's your turn and you can't move it's game over you lost I think this game is super interesting from a game design perspective with a lot of things I like and a couple of things I dislike but more on that later okay so I was playing this game with my girlfriend and she was beating me all of the time that really stinked so what does a man do uh of course I tried to find a strategy guide there was none so what do I do next okay I just write an unbeatable AI that teaches me how to play um actually it was a great

opportunity to write my first minmax AI I've never done that before but I know that it's not all the tricky and actually it wasn't so I'm going to explain to you how it's done how it works in hpow it's very simple to evaluate board positions because we simply have to count the stones on every side at the moment the AI has two stones less than the other player so that results in a score of minus two the goal of the AI is obviously to maximize its score value so how can I do that the easiest way to do that for the AI is simply to try out all of the moves that are currently possible and see which one results in the highest score so hey we're done this is the best move Yi um okay one second because we also need to consider that after that it will be the turn of the enemy and what if the enemy can instantly steal all of those stones back huh maybe in this case this isn't the best move after all the AI should go with this or this move instead because yes it doesn't gain any stones in the short run but it also prevents the

opponent from gaining any Stones remember we want to maximize our score so if we look at that bottom row those two other turns with a maximum value now the question is just how can we predict what the enemy will do how do we know this will result in a minus 2 or- 16 how do we know that and the answer is once again simply trying every possible move from that specific Bard position and assuming that the enemy will pick the best move which is obviously the worst move for us because it's a zero sum game but even though this looks like the worst move for us at at the moment we also need to consider the fact that after that it will be our turn again and we might be able to steal all of those stones back again the way we know if we can do that is simply by once again checking all of the options we have from that b position as you can see here this tree will literally keep expanding into millions and millions of board positions we need to check so that means we cannot go infinitely deep we need to stop somewhere because we are limited by processing power what we end up with is an insanely complex tree of options with

some evaluation numbers at the end currently it's the turn of white and white tries to maximize the final score value while red tries to minimize the final score value that's why it's called the minmax algorithm woe mind blown and now I have a little bit of a puzzle for you what is the best move for white just pause the video and try to figure it out and then tell me in the comments if you found the correct solution the solution is to go through the tree three backwards we know that red will try to minimize the score value so we can resolve that tree basically by picking the minimum value for every decision red has white tries to maximize the score value so let's pick the maximum value for every decision white has now red tries to minimize again and now it should be very easy to see what the best move is all of that is exactly what my hus AI does so now that you understand how it works let's have a look at it here you can see it calculating through millions of possible moves I lighting and green what it believes the best move

to be at the moment the AI looks seven moves into the future and it thinks it will have on eight point Advantage after that but if we let it think even longer way way longer we see that it changes its mind because now it's looking eight moves into the future but what happens if you wait even longer 3 minutes later there and now thei Chang mind again to this move because it's looking nine moves into the future now let's say there are about six possible moves a player could make every turn which feels pretty fair then that means the AI has already checked more than 10 million possible moves so imagine a tree like this but with 10 million numbers at the end so the more moves we calculate into the future the better our move is going to be but it will also take exponentially longer to compute calculating 10 moves into the

future would probably take something like 20 or 30 minutes I didn't have the patience for that so far so according to the AI this is the best opening move in huspa if you ever play this game you know what to do now isn't that great and then according to the AI this seems to be their best response and then for the second move of the starting player you should go for this and now here's where it gets interesting because I bet most human players would grab those three stones now to steal all of these Stones right here they can all be stolen but the AI decides not to do that instead it goes for this move which doesn't steal any stones at all the reason why it does that is because it knows in the long run this will actually result in a higher score as we know a move that looks good at first can actually turn out to be a horrible move so what the AI is probably doing here is damage control because when we have two AIS battling it out to the best of their ability the one that goes first seems to have quite the advantage it feels like the other one is more or less just trying to hang in

there as long as possible now you might wonder what happens when a human plays against something that can look eight moves into the future and the answer is the human loses every single time there are some games where looking eight moves into the future is actually not nearly enough to play at a superhuman level chess for example a good player can easily think eight moves ahead in h bow however it's difficult enough to think one move ahead because every single move you make can have completely weird Ripple effects and the entire board changes basically so eight moves for H bow is plenty plenty maybe you've heard about Alpha go by Deep Mind which is a go AI that destroyed one of the world's best go player Le it all and go is an infinitely more complex game I mean the rules are very simple but every single time it's your turn there are about 400 possible moves you could make so just calculating three moves ahead and go results in 60 million board positions and that means

three stones for a professional go player that is nothing so how did deep mine do it how did they beat a professional go player and surprisingly the answer is minmax they actually used minmax however they did not check every 400 possible moves every single turn they had a machine learning AI that can find the four or five most promising moves and they only check those this way the tree gets a lot smaller and more manageable and they can go a lot deeper also they use machine learning to evaluate the board position at the end so these numbers here at the end of the tree they are calculated by an AI as well so if you haven't wrapped your mind around it machines can already beat humans at every board game that ever existed we are not safe so yeah surprisingly the result was actually a superhuman hosb AI the cool thing about this is once we've got this AI set up it is very easy to change the rule sets and it will still play the game at a

superhuman level it doesn't need to relearn anything you change two lines of code and plays a different game at a superhuman level that's super cool that allows me to experiment with the rules you can use the AI to sort of try to get rid of the first move Advantage you can you can try to balance the starting positions of the stones out for example the disadvantage of an AI that plays at a superhuman level is of course that it is not representative of human play that means the game might now be balanced for super human AI players but just for casual Normy players maybe it's completely unbalanced or it was already completely balanced before so that is a downside that I clearly see that these superhuman AIS do not play like humans and therefore maybe they should not be used to balance games that are meant to be played by humans I want to keep the things I like about hospa unchanged and that is the unpredictability of the game the score tracking that basically the stones on your side also equal your

score it is a very elegent solution in my opinion that no Stones need to be removed from the game I do like how impactful every single move can be and the crazy Wombo combos you can pull off M and obviously just handling the stones feels so great I think we might never achieve that in video games it's just so nice to touch them what I think prod use some improvements is the fact that it's a little hard to do a comeback if you have fewer Stones then it's also harder to steal enemy Stones so there's basically a positive feedback loop that helps the winning player and on top of that whoever moves first also has a slight Advantage as we know now so those are the two things I wanted to fix here is what I came up with so this is a slightly different starting position and secondly you're only allowed to steal enemy Stones now when you land on an empty field and that in theory should help the losing player instead of the winning player because obviously whoever is behind has more empty holes and now we just lean back and let the a I play test the game as it turns out though I

designed a game that when played perfectly never ends damn we could obviously resolve that issue by adding another rule that maybe removes stones from the game and thereby forces it to an end but I didn't want to do that because that's one of the things I liked about the original game that can go back and forth and that all of the stones stay in the game so that was not a better game and I eventually decided that the original game as it is is already the best version so far obviously I kept experimenting I didn't give up that easily but nonetheless I didn't find anything that worked better than the original I also tried to get rid of the first move Advantage by trying a couple of different starting positions for the stones going to call that a fail as well but nonetheless I validated this as a very valid method for game design in my opinion this feels super useful it's just so nice to design a game and then instant ly see somebody play the game at

a superhuman level that is undeniably a very useful thing for game designers we are so extremely close to a world where there's not a single game where humans can actually beat machines just yesterday I watched Deep Mind beating professional Starcraft players 10 out of 11 games in Starcraft like which game is even left at this point and don't get me wrong AI has a long way still to go humans are still way Superior in many many points but when it comes to playing games like I give it three more years four more years and after that machines will just be unbeatable Untouchable in every game I'd imagine that earlier or later this will have quite a heavy impact on game design and on how we design games might still be a little early but let's better get ready for it I've got to say I'm so excited for the future how cool would that be just have an AI play test your game at a human or at a super human level and tell you if it's balanced and it finds all of your

bucks and exploits that would be so great if you think this topic is interesting then check out the other videos I've made about this topic I already tried writing an AI for a video game as well uh one the teachers itself actually another video I explain a bit how uh neuron networks artificial intelligence and all of that works and what it could mean for video games so if you're interested check it out thanks for watching have an amazing day see you [Music]