Thank you. Thanks for the introduction. So, hey, I'm thrilled to be here. Um, and thanks to the organizers, Fabian and Veronica. I think they're doing fantastic job for 10 years organizing this event. I think it's really good. I feel coming to a community. I live um in Stockholm in in Sweden and over the years I've really bonded with many of you. I see some of you also in the audience. So, it's really a fantastic uh fantastic opportunity to be here in Berlin. Um um and I am also German, one of my nationalities. So I'm always happy to come back to uh to Germany. But hey um I think the emphasis of my uh presentation will be going from single agent systems to multi-agent systems.
It's actually an area I'm very passionate about because I have been doing my PhD on multi-agent systems. I think many of you are now quite familiar with the first wave machine learning wave that we had for 10 years from 2010 to 2020 took basically all the air in the AI space and then we have had LLMs which are basically word models right uh so these are really quite narrow technologies in the in the AI world one area I will talk about now is multi- aent systems so it's not just some of you might think of agents as being LLM rappers So entities that you can talk to, you give it a name, a role and off you go.
Uh in terms of deploying them perhaps even in enterprises. I think the uh purpose of this talk will be more on um providing you a broader context and reframing that sense of what a multi- aent system actually is and the decadel long research that went into uh multi- aent systems um and what that actually means and why it's quite a different type of industry why it's quite a different type of um opportunity in terms of making some real business. So the first thing I would like to mention is um you can run we had some technical things there before you can run the video in the background. Um it's the AI ground challenge. It's the first work that I've been doing in the end of the '9s.
Um many of you are familiar with the touring test by now I'm assuming which some say hey we have passed it because we have and if you could just run the Are you do? Yeah. Okay. Uh and uh the touring test is about fooling a human and thinking well are you an AI or are you an actual human and some would say yep we sort of pass that. Uh the real challenges and I was exposed to that in 99. I wrote my thesis on that and worked with researchers around the world is really um the grand challenge of AI one of many but this is the one where you have agents where have AI set up playing soccer against each other.
I'm surely some of you have seen the videos and the footing the footage of these types of AI. So they are fully autonomous. These are agents. And the reason why this sort of domain is very challenging from an AI perspective, it's because it's fundamentally different than something like uh chess or alpho, right? So uh the big difference, right? If you like which has not so much to do with agent systems in the first instance is these are uh robots that are in a dynamic real-time world, right? Yeah. And um uh unlike Alph Go or uh in chess where you have a static field, right? So uh this started in 97. It was started by Kitano Hiro Hiroaki Kitano Manuela in the US.
Uh stating this is a grand challenge and I will just basically repeat what's on the slide. Uh it's by the middle of the 21st century a team of fully autonomous humanoid robotic soccer players shall win a soccer game uh complying with the official rules of FIFA against the winner of the most recent World Cup. All right. So the idea would be that these are maturing let's say in their development and they have a full ability to win against a human best team right by 2050. So obviously we are about 25 years away. If you see the robots that I saw in 99, this is a great advancement. We're now sort of halfway. I see the movement being very quick.
And what I want to point out here is that these agents, each of these robots, they have their own brain. They have their own perception. They need to have the ability to understand what it means to play in a team. They need to have an understanding of what it means to play against the team. So you have all these uh typical features that you would have in a multi- aent system, right? So you cannot this is not single agent. If you were to build this as single agent systems, you would not obviously not succeed, right? So um yes and you can go back to the presentation please. So yeah from a business side you know this has tremendous applications right obviously if you solve these types I mean just to wind back I live in Stockholm so we have the great pleasure and privilege to have all the Nobel Prize folks coming over when they got their prize and uh you know the one that was starting deep mind uh came over of course and
that's alpha fold and alpha fold go uh which is used to in deep mind to look at all the molecular um chemist chemical molecules right and that's a sort of still an an old style AI um but it has created tremendous amount of business right and we are all here and I'm assuming many of you are now having running businesses you know which are using AI in different forms so this is also creating uh business of course right this is not futuristic this is obviously used also in universities and for kids and so on but the real impact and the reason why Sony Honda many uh companies are putting a lot of money into is because it has this tremendous potential right you can use them on factory floors.
Uh Gartner has estimated that 50% of the um uh robots used on factory floors will use multi- Asian orchestration. So the way they communicate with each other and naturally you don't want to have uh robots which are from the same supplier so that they all homogeneous but you would have them for many suppliers which means that you uh need to think about how do you make them talk to each other, how will they agree with each other and so on and so on. So um I just want to mention this. I myself wrote my thesis and I got an award by the Swedish AI society back in the day. So it really spurred my enthusiasm to do more in AI, right?
Just realizing that I just scratched the surface of this entire domain 25 years ago. Uh another uh project that really got me quite busy was during my um research in Australia where we looked at adherence agents. They are used we use them um in chronic disease management and also in ser for service contracts for the British telecom. What you're seeing on the left is the original uh of some of the papers we have published and there's several companies that came out of it in Australia. Basically the idea here is that you have um a patient with a chronic disease diabetes for example and they are usually on very complex care plans right I mean if you have diabetes or cancer you have to go to the podiatrist you have to check your feet you have to get education is really complex and the the adherence uh on those care plans is very low most patients don't follow it for many reasons um sometimes they're 80 years old you know the complexity of these plans pro
and they are not and and what we know is if they were adhered to all these patients would be in a fantastic position right they would be very healthy because they are following up on all the on all the different um tests and things they should do u but they don't and so what we did we built an agent system uh to help all these caretakers and the care takers in this picture the podiatrist the general practitioner optometrists and so on to follow um their plans they're part of the plan. So we advise them on when they might have forgotten something. We noticed that you know the plan goes off track and we would provide information so that the plan would go on track again.
Right? So that's that's one of those things. We used BDI agents. They believe desire intention agents uh which is one of the most dominant agent forms that have been investigated over the last 30 years in the agent and multi-agent system community. They were used also by my supervisor for NASA to to plan the space shuttle rockets in the end 80s already. So it's a very fascinating um approach to build agent systems. So obviously several domains we use them also for with the British telecom for customer life cycle management. similar situations there. You have a telecommunication provider, you have any any uh uh customer and there are many factors that influence when and how the customer will buy more if they stay on track and you have uh you want to know when those moments occur, right?
When the plan goes off track so to say. And of course in healthcare chronic disease management, we had a couple of care pilots over in Australia. Uh jumping again, I went to Brazil and worked at the uh University of uh Rio Grande Desol in the south of Brazil. Great university. Stayed there for a couple of months and worked with Anna Bazan, professor who did a PhD in Germany and started some companies down there and she is the world expert in uh using agents for uh urban mobility and basically traffic light coordination. So it turns out that if you would like to have a um system working and you want to uh optimize the traffic lights, let's say in Berlin uh and you would assign every traffic light an agent and it should act optimally according to having um traffic flow being optimal and minimizing uh exhaust and pollution.
Right? So um it's an extremely complex problem. So if you have more than 20 traffic lights which surely most any average small city has the combinations are astronomical. uh so it's a very very a very complex problem and you cannot find a solution by brute force calculation right so uh one of the applications that we looked at is um reinforcement learning so I don't go into details ask me later if you would like to know more details but basically every traffic light would be represented as a tiny agent learner observing its space around it and it gets feedback and it gets reinforcement uh as to whether it turned the at the right time with a you know at the right timing and it would get a reward if the traffic flows faster or a negative reward if it doesn't flow so fast.
So the trick is here that every traffic light as an agent learns and um they can also communicate with each other. So they would say hey you know there's a platoon of cars coming your way just so you're aware so so you know what's going to happen right? So this is in fact last year the Wall Street Journal uh published uh some results. It wasn't by Anna Bazan but by Google showing that it has some real impact. I think it was used in the city of New York and there was 30% reduction of traffic in the area that they tested and 10% produ reduction in um pollution. So quite some very interesting things. Then um I fell in love with this whole idea of simulating entire societies.
So uh I went to Sweden when I lived in Australia. I came to Sweden. And I got an award by the Australian Academia of Science and worked together with six and the Carolin institute which is a a large university medical university in the in the Nordics in Stockholm and we looked at m so-called micro simulations which is not an agent system per se but what what we have done is we have represented every single Swedish person um as an agent and looked you know gave them certain properties as to where they lived, how often they would travel uh and all these types of things, right? And then in two it was already in 2000 2001 uh we simulated things like pandemic.
So well before corona of course like 20 years before corona and how this would influence behaviors and how much time you would have to uh build to create vaccines and how quickly you can sort of uh give them to the population. Right? So an additional thing that I looked at particularly was not just having micro simulations which are somewhat passive but I added a layer of intentional BDI agents. So agents that had a intentional stance meaning they have an ability to make decisions. In this case I looked at clinics I looked at uh entities authorities which made decisions based on the micro simulations output. So um that's that's another way of running agents and this uh work has also been used uh in different domains healthcare life science telecom obviously especially micro simulations for pandemics and healthcare have a big influence on the way you are organizing your resources within a country.
It's also used for EV chargers. So now with the um you know electric cars coming out we have a uh we we we want to know when and how people are charging their cars. So this is also another application domain of micro simulations. Agents really add some extra value to this entire picture. So yeah so these are some of the things I have been working on. I wonder if I forgotten anything important. I mean there's obviously so many more applications of multi- aent systems over the last 40 years. Automatic trading is one of those for example or algorithmic trading. So there you look at multi- aent systems which are competitive. So you use an agent or several agents.
You have to make an assessment of the other groups. Um and you have to do that really quick. So how quickly can you assess for example that other parties in the world are selling their stocks or their their trades. Right? So this is actually to Santom who has been building up I don't know how much his business is worth now half a billion or billion or something over the last couple of years which is a pretty interesting domain. So I turned this presentation around a little bit. I wanted to start with the applications. You get a real sense of how you use multi- aent systems. Uh and now I'm going back and summarize a little bit.
You know what what is an multi- aent system now exactly or not exactly but you know so you get a gist. And it's basically a collection of cooperating and competing agents that have either joint goals or individual goals. And um they have an assessment about their world. They have their own private assessment about the world just as any of us in this room have um and that is something that's private to you. So all the other agents, no other system has access to that type of information. Uh the agents must have therefore an ability to plan ideally collaboratively. There's a lot of great research done by gross from Harvard on having shared plans much of that is implemented in the real world.
Um and uh it means for example um in even in teams in rescue teams say where you might assume that you have full access or you work together you still need to have the ability of building a model of the others in your team their capabilities their intentions. Imagine a bushfire which we had plenty of in Australia and you have not enough communication. You need to understand the uh capabilities of the other colleagues in the field so to say. What cars do they have? Have they broken down? Is there communication that has failed right? So you need to have that active model about others in this in this field and I will actually mention a bit more about it.
The other very important part is this game theoretic concept. par to optimality nashik equilibrium many many of you have heard about it which means that if you're an adversarial uh environment and you need to compete like in Robocop or in algorithmic trading or if you start using agents um and deploy them in enterprises which I've heard many times using LLM wrapped agents right uh the most naive way by the way to do that is to build these LLM wrapped agents and let them on the loose you say oh you're a customer agent you are that agent you give it a little prompt and off you go that is almost doomed to be a failure, right?
you need to have all these other layers of organizational um protocols in order to make these agents behave in a way that you know you really intend them to behave and of course human AI teams right we speak of the system uh that you know the the agents that do their own thing but I think the key lies also in having these human AI systems which is one of those things that um I want to show at least one slide uh which is a little deeper from the technological point of view I hope you know I know it's late in the day but basically that's um part part of my thesis my PhD thesis back in the days and I I I am happy to show some of that after so many years but basically you have different agents on the left side and you see a mathematical representation as to how well teams or team constellations on the right side would actually play in the real world.
So you have a model as all of us have about groups about individuals and their capabilities and that's just a model it's not reality. So what you see on the right side is something that you do not know accurately which is well how well does that team actually perform in the real world and as you're seeing how they perform the models of these agents on the left side are updated and so that's that's one area which is extremely important because you need to understand you need to have models you never have accurate information about others. So that's really a function that you need to uh that you need to uh understand and give these agents as a capability.
So I think I I um it's okay. So let's go through the next few slides. So some learnings and these are by no means exhaustive or completely applicable and I would be more than happy to share more share more but it's like build building multi- aent systems for industry purposes right as what are the dos and the don'ts one thing that I'm seeing uh letting starting with the dos um in successful projects you really follow best practices so what I mentioned earlier you know building naive agents and have them run at scale you know just an LLM rapper is doomed to fail. And I've been advising a couple of companies, CEOs and VPs responsible for this field to really move away from that simple implementation because you just will fall on your face when you do that.
Um I mentioned some already but I want to point out just a few five of those in the middle which is like multi- aent system foundations you know task allocation very important contract net protocols a key cornerstone in the multi- aent system uh community have been deployed since the 80s I barely see them anywhere in use in real industrial applications especially when uh consultancy companies for example start moving into this space right um there's very few that truly have a sophisticated task allocation protocol in there. Um, agent modeling, I mentioned already a little bit about it. There are many many models of doing that. For example, there's a recursive modeling method. So, for example, if Mark and I, we talked earlier, right?
We if I was to have a model of you and look at your strategy as to how you would make a move in chess, right? let's say, and you have a model about me as to what my next move is. And we do this back and forth forever. That would be called recursive modeling. And we can do this until infinity. Of course, you cut this off in practice, but this is how you would build agents that are learning the um strategies that you that other agents would have. Yeah. So, and I mentioned earlier de reinforcement learning, decentralized planning, extremely important parts. Milin Tumblr, professor at Harvard, also a good colleague of mine, um is looking at something called flexible teamwork, where you look specifically at how you are making sure these uh how you take into account when communication fails, how you observing team members falling away and how does that change your plan?
Okay. And what not to do. Um I see multi- aent systems being implemented in practice and they that's a complete overkill. Like if agents I give you just some hints, right? But if agents have access to the same data and they have the same control and you start building a system where you give several entities autonomy is a complete overkill. You don't need to do that. If you can build one system and you don't need to worry about the overhead of communication and and coordination because all of these agents basically have access to the same things. No need, right? uh also this master slave architectures uh Michael Woolrage fantastic guy follow him uh on um YouTube or elsewhere he's professor from Cambridge and um he would describe an agent as an unlike an object or program as an entity that you could not just tell something.
So an object would just do it an agent would ask for money. So that's one way he would describe the ability of an agent to make autonomous decisions. Right? And of course you know all agents are not the same. This is sort of the open system assumption. The reality that we live in is not we are not living in a closed system meaning we have an enterprise or certain business area and the agents are all assumed to do the same thing. We have full excess of what they're doing. None of them will ever fail. This is too naive. Most of the time we have indeed agents that have uh a lot of limitations including bounded rationality and misbehaviors in many ways especially if they are if you build them in a way that they maximize certain functions.
So and then I want to just talk in the end about um an area that I find extremely fascinating. I've also been working on that for the last 25 years. But I see this area as being something that's going to emerge in the next say 5 to 10 years or so. I just spoke to Peter Sarland too um was earlier today in the panel. Unfortunately, I couldn't see it, but he's uh he in fact also invested in a quantum company and I think it is extremely exciting um technology which is starting to become mature uh and uh especially for some of the problems I mentioned earlier like for example the multi- aent system um reinforcement learning uh problems quantum computing really holds promise to solve these types of problems fast um I won't go too much into it I will just leave you with a thought here that this is something I'm moving in.
We have a pretty exciting spin-off um that we have now uh worked on from the Nordic AI Institute. That's the institute that I have founded um in Sweden and Stockholm. And one spin-off is indeed focusing a lot on quantum quantum AI and on um using it actually in the application domain of pharma. Uh yes and I might just leave it there for now so we have a bit of time for questions. Thanks. Thank you. Any questions from the audience? Maybe about quantum computing. All up for it. Oh, who's who's can I ask her? Who's into quantum? Who knows stuff about quantum computing? Okay. I moderated a conference two years ago at tomb technical university in Munich and there we talked about quantum and also about implementing um like using brain cells as computers.
Um but it was hard for the audience to grasp. Yeah. A good so good choice for me not having gone into too much detail is that you're saying well because it's such a different concept of Yeah. Yeah. It's a completely different concept. Yeah. this would deserve of course an entire day. I mean it's they're different topologies and it's not like the normal computing right with bits and all this I mean you have concepts like decoherence where the quantum state is falling away and uh meaning that your program disappears you have only a couple of cubits which really work but turns out there's a lot of work here which um is becoming very practical uh and it does work in very limited domains and um I I'm I'm quite bullish on quantum partially because I have been building some of those paper quantum circus on a paper myself, but seeing now how how it's moving forward is really exciting.
You have access to a quantum computer. We theoretically we all have I mean there's about 10 in Europe and you can certainly uh ask for excess and then go for it. Uh you should if you haven't done not sure would you ask for access if you're not a research institute. Ah you can you can go to several several uh platforms. Uh you can also go to for example I'm doing a little bit of uh advertisement here for young guts and IQM. It's probably the biggest uh quantum computer in Europe. Uh he's just now raising a funding round I think of 125 million and IQ they just released two weeks ago 50 quantum bit computer and that's accessible.
You can uh go to their website. You know there's Pascal in Paris. I was two weeks ago in Paris and met Pascal, Alis and Bob. Um there's obviously in Germany you live I think in Hamburg also. So there's a couple of people mostly know about you in Germany. But I do I was trying to chime in if I may. Yes, please. I'm sorry I didn't see you. Apologize. I'm a partner with the AI fund and also an old quantum physicist. I have done my PhD in quantum physics. Fantastic at the time on uh quantum states of the light field. Um I am less bullish on quantum computing as an investor at this point in time despite the fact I wrote my PhD thesis on it.
Um because I wonder about the required amount of cubits and the decoherence times versus the calculation times for practical applications meaning if you only have 50 bits yeah there's only so many real world problems that can be can be put into that small a solution space. Yeah. The massive parallelism works as long as the coherence doesn't kick in. So correct the soon that the system starts interacting with the world around it. Yes. The superp position starts to fade away and your your results get worse and worse and because they are also always probabilistic in nature the outcomes. Yeah. It sort of you get no real answer. Yeah. Therefore I am wondering whether it's the right time to invest now.
Yeah. I think it's an amazing time in terms of making breakthroughs, but my personal view would be that I would wait a little until sort of the number of cubits at a decent calculation time reaches 10,000 or so. Yeah. And before that then it becomes really applicable. Apart from I would make two exceptions. Yeah. Like quantum simulations. We've seen that quantum supremacy has worked in areas where it's quantum computers are used to model quantum systems. And the second area is if um you try to do cryptographic problems or you have a key of a certain length which isn't too long that you can crack effectively. So uh so yeah to put some water in the wine no I think you're talking about the the number of cubits you need.
Yeah. It's it's funny enough I have these discussions more often with quantum physicists who have done their PhD and they say I'm not so bullish but I don't know if you have followed you must have followed like Google's and and Microsoft's my I know it's like you know boosted out to this year by the way is the year of the quantum year 100 years ago shringer went up to Helgo London wrote down his equations right 23 years old important that's 100 years ago was 100 years ago yeah when this whole you know spooky interactions that Einstein and others called was a I'm just always thinking of Max plung in 43 when he gave the big speech about quantum physics that a lot of things that he um developed or explored uh he cannot answer these questions with physics and leaves a lot of questions for the philosophers now.
Yeah. Yeah. But things have changed 43 43 I wasn't what would he think now? What would Max plank say now about our established like research in quantum physics? I think you would be surprised how how far we have come because about 20 30 years everything we have now would have been considered be science fiction right like but we have come far the fact that we have cubits at work is good and I would say I would I would push against what you mentioned in a sense that this is uh possibly pointing to a little problem now in Europe which we all say oh we are a little slow in our investment and so on right I would say this might be a wrong attitude I know how you invest in AI and it's good but this is the time now to actually take a higher risk because I'm optimistic that this stuff will come out.
Maybe it takes 5 years, 10 years, 20 years, right? But now is the time. Yes, the uncertainty is higher. It's not like we know LLM works and all this works already. We are not at the forefront of the existing AI technology because we were too risk averse 20 years ago. And so now in my opinion uh this is an excellent time and I'm you know these many researchers you know IQM Pascal in in Europe we have fantastic researchers uh we have also the will to invest you know Peter Salen is just here on this conference he invest this is a important time to not miss the train and we have already 10 times less investment in quantum here in Europe as we have in the US so let's not miss that train that's my you know even if yes lots of questions I agree no running quantum computer but let's not miss the train.
Well, I would say that's my message. That's a good message to end your presentation. Thank you so much Dr. Christian. Up.