Hello and welcome. I will be talking to you about AI in the fast lane, how we can use smart traffic systems to unclog our cities. This way we can use intelligent technology to revolutionize urban transport. First couple of things about me. My name is Ara Del. I have been working as an engineering manager for the last seven years. I have worked in various technical roles for the last 13 years in the automotive, FinTech and tech industries. I have worked four years in automated driving and I have an electrical computer engineering degree a PhD in the field with focus on multi objective optimization. And I have a degree in education. Throughout my career, I have been mentoring and trying to grow professionally, computer engineers. And fun fact, before switching to the industry, I was working as a teacher. I was teaching for seven years. But let's yeah, let's look into. More detail and what we will be discussing today. First, I will touch a bit on what the current challenges are in traffic management systems. Then I will discuss a bit how we can use software agents and smart traffic systems. Then I will touch briefly on what the required infrastructure would be to be able to implement. So it's a system. And finally, I will present what the benefits can be of using ai driven traffic management. Let's get started then. So what are the challenges? I'm sure everyone has faced, if not all of them, most of them, and everyone is familiar with 'em. First we do have an increase in congestion. Yeah, congestion is rising in urban areas. We have cities being overpopulated, population is growing, and this also leads to the increase in vehicle usages, and this leads to longer commute times and congestion, the infrastructure in our cities. It is often inadequate. This can happen either because it's outdated and it hasn't been maintained for some time, or because it was not built to withstand all the pressure it feels now with all the vehicles and all the road users that increase in number. I already mentioned that we have a growing number of vehicles. This doesn't happen with the same rate in every country in the world, or every city in the world. But the trend is still increasing and this poses significant challenges to the traffic management systems. This all this also generates the other problem when we have traffic events. Usually our reaction is delayed. Our systems are too slow. Usually some human needs to detect something and some human needs to intervene, and that can take a lot of time. So the result is lost time, lost fuel, or other type of cost, and most importantly, lost lives. But what if. We didn't have to have such a traffic management system. What if every car and traffic light was an intelligent agent? What if they could negotiate, adapt, and optimize their behavior in real time? That is the idea behind the decentralized traffic intelligence. Let's, I'll look into it a little bit more. How would we use software agents? In a smart traffic system? There would be mobile agents that would correspond to the vehicles, and then in the intersections we would have stationary agents. Each of them makes decisions based on local context mostly. However, if we feel that there's a need for that, we can use more centralized optimizations. And more centralized policies so that we can try to optimize for the broader urban area not just locally. And this is the, a more system which stands for adaptive multi objective, optimized routing agents, optimize for time, distance, cost emissions, or any other objective that one might thing makes sense. The agents bid for road segments based on current values of these objectives. And the routes evolve continuously as the traffic changes. So in more detail, a vehicle approaches a decision point, like an intersection. It requests a potential next step from the nearby stationary agents. Each stationary agent. Calculates a bit a cost estimate for using the road segments over route. It proposes for the vehicle to take based on the current data. Then the vehicle evaluates all bids using its weighted objective function and its selects the lowest cost next step, which is the best trade off among the priorities we have selected to optimize by. How is the bid negotiation happening? So the vehicle agent sends the route request to the nearby stationary agents, plus the wage for each objective to optimize the stationary agent. Then accesses segment data for each objective. And depending on the data freshness or staleness, it may calculate or update expected values. For the segments of the potential routes for the different objectives that we want to optimize by then the stationary agent will normalize the values for each of these segments. It'll calculate the cost per segment using the weighted objective function, which in the case of three objectives that are time, distance, and monetary cost. The cost of the segment would be the weight. All the time, multiplied by the normalized time, the weight of the distance multiplied by the normalized distance and the weight of the cost multiplied by the normalized monetary cost. So the stationary agent will calculate the possible routes and that the pico can take. As mentioned, the cost of each route would be the sum of the cost of each segments. The stationary agent bid will be the cost of the lowest route cost, and this will be the route that it would be recommending to the vehicle. So the vehicle selects and follows the initial segments of the lowest bid route. Then the process continues with re-planning as the vehicle moves. The vehicle reevaluates the options, every few steps. Why this happens because conditions can change. Like we can have congestion present that wasn't there before. We can have an accident or some other type of emergency can occur, so the stationary agents at the next step or some steps forward could potentially suggest better alternatives based on the new conditions. And this allows for real time rerouting and self adaptation. Let's see how the big generation process would work for yeah, three objectives monitored. So we would require a route, we would query the database for the network segment data, we would get the segments for. For the different routes that we could take. We would estimate the expected traffic vehicle speed for these segments of the road network. And then we would be creating our network graph using the distance, time, and cost in this case where we have these three objectives. Yeah. And the waves would be according to the profile provided. We can use some algorithm to figure out which one of the potential routes would be the optimal. It would be Dextra or it could be some other more sophisticated potentially. And then we would see if a route exists, then we would. Calculate the bid and send it to the vehicle. If it doesn't exist, then we would send some negative value and then the vehicle would select the lowest positive value that we could get as a bid. And let's see what happens next. So we said a route exists. If it doesn't exist, we cannot go to the place we wanted. So there's nothing to do. If route does exist, then we receive a list of the next nodes that we need to go through and then we start our journey and we see is the next road congested or is the list. All the next nodes complete after this one. If not, then we move to the next node and then we check if the next node is our target node or not. If it is a target node or it's our destination, yay, everyone is happy. If not, then we move to the next node after that, and then we check again. Is our node list complete? Have we been through all of them, or is there some type of emergency or congestion in the next node? If there is some emergency or congestion or the list is not complete, we initiate an auction again, and then the process continues until we finally reach our destination. Here's a conceptual diagram of how this would work. You can see vehicle agents, you can see different type of mobile agents like the vehicles, emergency vehicles, public transport, and so on. And you also see the stationary agents. We do collect real time traffic data, and it can be that stationary agents can also exchange data. In case it will improve their estimations for the values that they will provide for the different objectives. So what I did was tried to run some simulations to see how effective such a system would be. So I used the net logo simulation environment and I run lots of simulations. A few hundreds with different scenarios that included cases where there was congestion, emergencies, different type of vehicles and so on. And this happened for some fixed distance. And this is the results I got. So for the same distance, I compared three different. Approaches. One is this normal routing distance base. You try to find the route with a minimum distance and then you execute that same thing. You try to execute the route, you find this time using multi objective routing so you can optimize for different objectives. But once you, decide on some route initially, then you execute it to the end. And then the third approach is the amor approach where you do optimize for different objectives, but you do dynamic re-planning every sum nodes or as discussed when there is some type of change, like an emergency or congestion that wasn't present before. What we see is that. The cost, monetary cost, when you have the multi objective approaches, the dynamic and non-dynamic re-planning is more or less the same, but it is lower than planning route based on distance only. And then we have a shorter travel time, using multi objective optimization than we would have optimizing based on distance. And we have even shorter duration of the travel when we do the dynamic re-planning with the amor system. And we can finally see that there's really a huge difference in how much. Our trip duration would increase if we suddenly get consistent in the traffic. So there is some slight difference between the distance based approach and the multi objective approach, but a more with dynamic or planning is really shining there with less than 20% of the increase compared to the other methods. And that's all very nice. What would we need to actually realize this in our roads? First, we would need V two Xen enabled infrastructure. For those that are not familiar with V two X, it means vehicle to anything and it allows vehicles to communicate with each other with different traffic elements like smart traffic lights. Or smart traffic signs, which can adapt what they display. And even with mobile phones potentially so the vehicle could communicate with the pedestrian that comes behind the building in the next block. So you could really see and understand how the environment around you is. Besides the smart traffic lights and the adaptive road signs, we would also have the roadside units, which would be the stationary agents. So from the vehicle side. Vehicles would obviously need to be V two X capable and the vehicles would be running the actual app. The agents that they would use in order to to their smart road planning and communicate with the rest of the infrastructure. On the cloud and edge layer level we would need local road terminals. This would enable a fast computation and assuming we want to have some centralized overview of the traffic and potentially use this to do a more centralized planning, we would need a city traffic management center. What assumptions are we making? So we're making the assumption that the map of the environment is known. So the vehicles do not do continuous, an online localization and mapping. They know the map of their environment already. The road segments may have tolls or different charges. We do have available historical traffic data and we continuously collect as the system is online, additional ones and the vehicles know their global position. Okay. We talked about everything, but what's the benefit? What would we gain out of this? We have seen a few things based on the simulation results, and this allows us to to estimate that this would make it easier to optimize traffic flow B two X enabled traffic. Signs and lights also play a role in this because you can change the signal timings based on the traffic and reduce the delays. AI driven systems that analyze in real time the traffic data play an important role in enabling this. This would also help us reduce bottlenecks. Both because we would try to optimize traffic flow, but we would also potentially redirect traffic and through different routes that would enable them to have a smoother journey and less congestion. And all of this would lead to sort travel times which in our current urban, our current environment in our urban cities would make a big difference for a lot of people that have to travel for a significant amount of time, spend some time on the road they could spin it in something potentially much more pleasant and productive. Besides this, using the ai driven traffic management system, we could have improvements with respect to road safety or how we prioritize our traffic. For example, when we have real time yeah, road closers something happens, we can rear out our traffic. Smart systems will enable us. To do this in a faster way and this will optimize traffic flow. And this would also help reduce potential accidents when there is a road blockage. We could also manage traffic in a priority based way so we can dynamically prioritize different actors and like public transferred or emergency services. And in the case of an emergency, we could even help ambulances negotiate some green wave access. So they get they get priority and they can go fast where they need to be. And finally, there are also environmental benefits and reduced emissions. We have a smoother traffic flow. This leads to reduced congestion but also we don't have this stop and go traffic all the time as the traffic is smoother. And this enables vehicles to consume less energy. Emit less CO2. And this also improves the air quality in our cities. Thank you very much. Please share also your thoughts, your views, or how you envision traffic management in the future. Thank you.