 I am going to talk about the artificial intelligence research in games, maybe a slight correction my thesis will not the first that studied AI in games, but it was probably the first that studied AI in video games, which is a topic that nowadays gets quite a lot of attention but it didn't when I wrote my thesis. When you talk about games you might wonder what is a game and to show you very quickly the two attributes that I think are essential for games is interactivity, you can act in a game, and the second is that it allows for consequence free experimentation that is playing. And that might make games very interesting as a study topic in many different domains. I look mainly at AI in games. AI and games have a very long history together when artificial intelligence was first coined as a term. The game of chess was immediately defined as a challenge for artificial intelligence, a challenge that was bested in 1997 when the IBMs the blue defeated Gary Kasparov. Until that time, most AI in games research was focused on these two player deterministic board games such as chess. In the turn of the century onwards, other games also started becoming a research focus such as video games, modern board games which you play with more than two players which have imperfect information, and also things like that fairly recent table top role playing and when I introduced this topic of game research to my students I usually show them this picture, which shows a game as a content, let's say the game world and agents that is the players of the game, and a player who interacts with the game so the ages might be AI players. And around this you see lots of different topics of research that was for instance player modeling that's something that I focus a lot on. That's also how you build a game that could be with an assistance of an AI, or it could be with automatic game creation with for instance procedural content generation, or if you want to generate a story computational narratives. So you can also see the topic of search and planning that's the typical three search techniques that I used for these two player deterministic board games, and the agents they bring a lot of research with them, namely general game AI that is artificial intelligence that is a game behavior learning that would be an agent that imitates a human and believable agents that would be agents that you would mistake for a human. And for these techniques, these topics, lots of different techniques and I just plopped a few down here I'm not going to go to them but you can see that many different AI techniques can be used for building game AI and for studying game AI. Now in recent years and I've already came up with the deep convolutional neural networks in recent years these networks got a lot of attention in game research, and that's mainly because of Google DeepMind, which did quite a lot of work since 2015, which they, at least published since 2015, on the use of convolutional neural networks in games. And I want to quickly go through these, and tell you very little bit about them, and also give a bit of criticism on them or at least remark, make some remarks on them. So the first is Atari games that was in 2015 Atari games are fairly simple games as space invaders which you can see here. And what they did is and that is what typically what convolutional neural networks I could for they loaded the screens into the neural network, and then let the new network learn how to play these games and they used a bunch of games but for about 27 of them. They managed to create a eyes this way that were super human that played the game much better than humans would, which was very impressive and I was impressed at the time, especially since they just use the screenshots, but I was also thinking isn't this overkill. And yes it is overkill, because there's a paper published recently, which is research that is slightly older but and when I heard about it and I thought yeah this is a revelation, revelation sorry. These games also with neural networks but each neural network had no hidden layers and only six to 18 neurons in them. And the same games could be played with this neural network, except that not all of them were equally good as as deep might achieve, although for some they played the games even better. So it indeed shows that you don't need such deep convolutional neural networks for this. Now what is incredibly impressive is what they did with the game of co which they managed to create an AI for called Alpha go that is defeated human champions in 2016. And this is probably the best approach for such a game, although it's not only that it uses deep convolutional neural networks that also uses a technique called Monte Carlo three searches really a combination of these two techniques. They do it with both training by human gameplay and with self play that is called Alpha zero. And this is definitely a very impressive achievement, except that I would like to remark that there is a huge difference between how Alpha go place go and how a human can do it. Because if you take Alpha go that place go on the standard 19 by 19 board if you would have a human grandmaster who plays go on that board in shape can you please pay it on a board of 17 by 17 human grandmaster will be able to do that immediately but Alpha go will not. It's like it's it's overfitting on one particular game, namely the game of go on the 19 by 19 board. If you think where you think if you want to study how humans play this even very good humans how they played. It's probably different than what Alpha go does. The last one that came up is Alpha star in here they diverged in the area of the video games they used to start craft which is a real time strategy game, and they trained and deep convolutional neural network to play. They developed on a grandmaster level. And they actually in January 2019 they defeated to starcraft grandmasters. Now, there is a lot to remark on this. First of all, the bot that they developed only can play one map. The second is it can play only one race in the game. It can only play on one particular version of the starcraft engine. There is access to information that humans do not have and finally, and this is the sneaky bit. What happens in these kind of AI is that they usually contain weaknesses and humans can exploit them. So to avoid humans exploiting the starcraft AI that they built, they actually build lots of different AIs, and every game they played against humans they gave them a different AI, because Alpha star is a collection of many different AIs. So the humans could never exploit these weaknesses. Now that's to be said, and you can also read in the paper is that they resolved some of these weaknesses in later versions. But at the final version, at least that was the version at the end of 2019 Alpha star is strong in playing starcraft, but not superhuman. If you look at the equivalent of centuries of training, it needs a lot of tweaking. It does no longer learn after it has been trained. It still has exploitable weaknesses and it's still limited to a specific version of the engine. So if you're thinking about artificial general game intelligence, this is a far cry from it. The little bit of an issue is that because of all the successes of deep minds, these deep convolutional networks became like a universal solution everybody wants to use them for everything. But are they a universal solution. I would say not and in the last minutes that I have I want to quickly go to a couple of games where these deep convolutional networks are not able or probably not able to solve the problems that these games offer. And this is still on the edge this still might be doable but deep convolutional neural networks is games often even higher complexity than I just discussed came like a Rima to the left has certain complexities which I will not go into. On the right is that they go you might know that they go the complexity of that they go doesn't seem very high, but it has a lot of imperfect information people have tried to train deep convolutional neural networks to play this game. And the resulting AI is at the moment still weak. So the techniques must be further developed for that. But I think it's achievable, even with these neural networks. Here's another issue games for more than two players. So the problem which you have here is that if humans play games which have for instance three players in them what you will see is that the strongest player will suddenly find that the other players are going to form an alliance against them. And this is something that is very hard to make a neural network for because you don't only have to find the best move, but also the move that will not entice the opponents to form an alliance against you. So this needs more social awareness. I still think that the neural networks might be part of a solution here, but it's definitely not a complete solution. The third is games with open ended action spaces, where for instance you have interactive fiction which you play by just typing English sentences and they should accept any English sentence. And that is a problem for neural networks because they usually have a limit action space, or to the right you see the role, tabletop role playing games that is humans playing a game around a table, writing a story together and having an AI be the game master or a player for a game like this requires an enormous amount of social awareness and understanding of the game world. So this is definitely beyond what a neural network can do. General gameplay is a topic like that in general game playing you just get a game description I have to play the game immediately. These are challenges for this they work in board games and a video games. Again, this is done usually with Monte Carlo three sets at the moment. And finally you have the whole idea of playing like a human. This is something that we actually don't know how to do with playing in such a way that you cannot distinguish the AI from a human. So what I would want to end with as a remark is the following. In games we find a wide open landscape of challenges for artificial intelligence for which we at present have no suitable technique of solving them. So if artificial and general intelligence, which is something that most AI researchers strive for should ever be in our reach. We should be able to take on these challenges in games because they are much easier to solve in games. And therefore I think that games as an application can be a crucial stepping stone in further AI research and development. Thank you.