INTEGRATED TRANSPORT SYSTEMS WITH ARTIFICIAL INTELLIGENCE CONTROL

72 | P a g e DOI: 10.5281/zenodo.17359016 “INTEGRATED TRANSPORT SYSTEMS WITH ARTIFICIAL INTELLIGENCE CONTROL” Mahadeva M1*, Vijet V Naik2, Abhijith C C3 1Assistant Professor, 2Undergraduate Students, Professor and Head3 Department of Civil Engineering, RNS Institute of Technology, Channasandra, Bengaluru, India *Corresponding author: [email protected] Abstract: City traffic gets more ridiculous every year. that never end, people driving like bumper cars, and don’t even get me started on the fumes. The old-school traffic lights those fixed timers or whatever “smart” sensors they claim to have just don’t keep up when everyone suddenly heads to the same brunch spot or a random rainy-day jam everything up. But here’s the wild part the wizards in tech have whipped up deep reinforcement learning (DRL picture computers learning when to make you sit or go) plus cars that talk to each other (CAVs for the acronym nerds). Basically, brains in the intersections. This review dives into how the newest DRL powered traffic lights try to juggle all three headaches at once speed, safety, and not choking on exhaust. the researchers are obsessed with things like D3QN, DQN (don’t worry about the alphabet soup), and these squad-like multi-agent setups, which all seem to cut down on wait times, pollution. Is there’s still a bunch to do like, weaving all those goals together without tripping over their own wires, and nobody’s totally nailed rolling this stuff out in actual, messy cities with real people yet. The paper flags a bunch of holes where research still needs to catch up and tosses out thoughts on how to make this tech work in the wild, not just on some fancy simulation. Keywords: Traffic Signal Control, Deep Reinforcement Learning, Connected and Automated Vehicles, MultiAgent Systems, Carbon Emission Reduction, Road Safety Introduction Ever sat stewing at a red light for what feels like half your life, you know traffic is a massive pain in city life not just annoying, but it's this black hole sucking up our time, our gas, and just belching more junk into the atmosphere. Longer waits, more chances for fender-benders, and pollution cranked to eleven. The old way city planners handle this is kind ancient, programmed schedules and uptight timing algorithms, basically hoping yesterday’s traffic patterns didn’t change. Enter Artificial intelligence, and more specifically, all that jazz with Deep Reinforcement Learning. It’s not just a buzzword this stuff “learns” by messing around in a simulation, picking up on traffic weirdness in real time. It’s like having a traffic cop that keeps levelling up, minus the whistle. Now with smart cars and connected vehicles zooming around, everything’s talking to everything else. Cars gossip, intersections eavesdrop, and the whole network just flows better at least, in theory. Most of these shiny systems only obsess over one thing. They either cut emissions, or speed up commutes, or try not to get people killed. 73 | P a g e DOI: 10.5281/zenodo.17359016 Rarely all three. But real cities. They need all of that, not just the greatest-hits version. It's digging through what’s already out there on DRL-powered traffic signals, calling out what’s missing, and pitching a new way to juggle all these priorities at once keep things quick, safe, and not turning the planet into an oven. Figure1: illustration of the scenario of the traffic signal connection Source: (Xiaoguang Yang and Jintao Lai 2023) Literature Review Deep Reinforcement Learning (DRL), particularly in the context of Connected and Automated Vehicles (CAVs) and Intelligent Traffic Signal Control (ITSC), to enhance traffic efficiency, safety, and environmental sustainability. The surveyed works generally agree that DRL methods offer significant improvements in reducing emissions, travel times, and collision risks compared to traditional systems. The discussion synthesizes key findings across multi-objective optimization, eco-driving mechanisms, and safety improvements achieved through various DRL algorithms like D3QN and DQN. However, the field continues to grapple with challenges related to scaling these solutions to massive real-world urban environments and ensuring stable reward function design. Wang and gang (2023) [1] basically tossed a D3QN model into the traffic mess that chaotic blend of robot-cars (CAVs) and good old human drivers and let it crunch the numbers. Cars magically spent less time puffing CO₂ and more time moving; bigger bonus once more CAVs showed up. Kang et al. (2024) [2] had a different party trick their so-called "Carbon Inclusion Mechanism." Sounds fancy. It’s basically handing out carbon credits for behaving nicely (eco-driving), and bam close to 18% cleaner emissions, plus folks didn’t have to slam on their brakes so much. Not ignored. Karbasi et al. (2024) [3] whipped up a DQN-based adaptive signal controller and no surprise rear-enders and dangerous crossings dipped, at least when you squint at the Time to 74 | P a g e DOI: 10.5281/zenodo.17359016 Collision stat. Meanwhile, bakhsh and Aziz (2024) [4] dropped a multi-objective D3QN-ATSC beast that cut down conflicts by 16%, trimmed emissions by 4%, and shaved almost a fifth off waiting times. Beats sitting at red lights staring at your phone. But wait, there’s more teamwork makes the dream work. Guo et al. (2023) [5] came up with COTV, a multi-agent, deep RL gizmo that lets signals and CAVs chat; this thing basically slashed both emissions and travel times by about 30%. Folks et al. (2023) [6] got creative too, mashing up local and global agents (MOMA-DDPG, if you’re keeping track), which wound up beating the pants off other DRL methods on throughput and eco vibes. Design always a spicy topic. Schumacher et al. (2023) [7] warned that squeezing DRL agents with straight-up CO₂ numbers is like feeding them hot sauce: unpredictable and not great for learning. Turns out, using proxies like queue length and how often folks must slam the brakes works way better. And back in Mascia et al. (2016) [8] (ancient history in AI years), showed you could knock out about 3% of nasty Black Carbon just by smartly routing and syncing signals. Talluri et al. (2025) [9] pushed the good old “Green Wave” lining up green lights so the whole street glides along, which means less unnecessary stops and a better deal for pedestrians and cyclists dodging traffic. Kim et al. (2019) [10] brought reinforcement learning into the shiny V2X world, tweaking signals in real time so everyone stops less and pollutes less. All this deep RL and CAV crossover action is moving the needle on efficiency, safety, and making the planet a tiny bit happier. But and it’s a big but there’s still the hard stuff like scaling this magic trick to massive, messy real-world cities, not to mention making sure reward systems don’t tank and, you know, getting out of simulation and onto real asphalt. Research Gap Lots of folks pick a lane they either chase lower emissions or hammer on safety but juggling both at once? Not happening much. Those fancy teamwork frameworks (COTV and MOMA-DDPG) Cool ideas, but shove them into real city traffic, especially with all sorts of cars and chaos, and they kind fall apart. It’s all gas-guzzlers as the default. When people use stuff like CO2 as the score, training the models goes haywire. Unstable as hell. And honestly, most of this research lives in make-believe land. It works in simulations (hello, SUMO and VISSIM), but you see it finding any actual large-scale pilots. single-objective focus, where most studies target either emissions or safety, failing to effectively juggle all three priorities efficiency, safety, and environmental impact simultaneously as required by urban environments. Real-world scalability and validation, as most research is confined to simulations, with actual large-scale city pilots being rare. The design of the system itself presents further challenges, notably the reward function instability when attempting to use direct CO2 metrics, leading to unpredictable results and a reliance on proxy metrics like queue length. Finally, current models often overlook the integration needs of Electric Vehicles (EVs) and fail to adequately address the critical issues of cybersecurity and the lack of standardized datasets and metrics necessary for objective algorithm comparison. Implementations Getting a multi-objective DRL-based Adaptive Traffic Signal Control (ATSC) thing off the ground it’s not as simple as flipping a few switches and calling it a day. Got a build this whole bridge from simulation nerd-land to the actual chaos of city streets. A solid simulation zone. Stuff like SUMO or VISSIM, they’re perfect for playing mad scientist with all types of cars old school gas guzzlers, shiny new electric rides, self-driving robot cars. EVs 75 | P a g e DOI: 10.5281/zenodo.17359016 got a get a special shout out. EV like regular cars they brake weird, need juice more often, and basically march to the beat of their own (carbon-neutral) drum. Think of them like video game copies of real intersections messes up way less when you screw things up mid-test. At the heart of this DRL-ATSC beast, you’ve got the reward function. Old models, they just cared about one thing maybe travel time, maybe emissions. Yawn. The new mantra is balance. That means traffic keeps moving, nobody’s playing bumper cars, and the environment’s not screaming in pain. We’re talking classic stuff like how long folks are stuck at the light, how many are lining up, plus the spicy safety bits TTC, those crap, that guy almost sideswiped moments, hard-braking freak-outs. Layer in green vibes too: CO2 numbers, stops, and how much juice or gas the whole circus burns through. Because if you chase just one goal, you always break something else. When it comes to the control side, you can’t just put one brain in charge of the whole city. That’s a meltdown waiting to happen. Nope, hybrid’s the way smaller agents for each little intersection (they’re the quick-react crew), then a big boss agent overseeing the lot, so traffic doesn’t get jammed two blocks down the road. Edge cloud split is where it’s at local edge stuff crunches real time moves, cloud brain handles the big picture, “where are we heading in five years” stuff. If you’re ever a run this show across a whole metropolis. People to “drive nice”. Got a something in front of them. Enter carbon incentive. Want drivers (especially those CAVs) to coast, stop hammering the throttle, and more like hand out some kind of credits or tokens for that behaviour. Tokens could mean cheaper tolls or charging perks. Suddenly, folks care. Slow and steady unless you like headlines about total gridlock. Start with a few “problem child” intersections, prove the thing doesn’t implode, then expand corridors, then citywide empire. Drag everyone to the table traffic bosses, car companies, city planners. Figure 2: Flow chart of implementation analysis 76 | P a g e DOI: 10.5281/zenodo.17359016 The project begins with two parallel branches GIS Route Optimization to identify the most cost-effective routes and Accident Analysis to identify and enhance high-risk areas known as "black spots" for increased road safety. These two distinct initiatives then work together to install sensor-based traffic signals to control flow in real time and evaluate the carbon footprint from increased efficiency. This results in Integrated Decision-Making, which combines knowledge from efficiency and safety. The project's development phase concludes when the combined plan is scaled up for full implementation after being tested in a pilot project. Future Scope Real adaptive traffic signal control using deep reinforcement learning (DRL) and connected autonomous vehicles (CAVs) is just getting started. It’s wild, and it’s gone a cover so much more than just cars idling First off, there’s the whole electric vehicle (EV) angle that’s really heating up. Imagine if your city’s traffic lights talked to the charging stations heck, maybe even to the grid itself. As more folks jump on the EV bandwagon, those traffic systems could help balance the rush for charging, even accounting for solar or wind energy when it’s available. So, you’re not just reducing traffic jams, you’re giving the planet a break, too. Then, about multimodal traffic right now, most of these fancy systems are obsessed with cars, while pedestrians, cyclists, and buses are basically left to fend for themselves. That’s got a change. If we want cities that don’t make walking or biking a near-death experience. Figure 3. Percentage Reduction in Emissions with green sustainability Source:( Kranthi Kumar Talluri2025 ) bake them right into the optimization game. Better reward systems and new control tricks should make cities more liveable, not just a little faster for SUVs. Tech wise, you’ve got edge computing, the cloud, and this thing called federated learning. That bit just means cities can help train smarter, more adaptable AI for traffic, all while keeping your data out of the greedy hands of central servers. Privacy, check. If one part of the network gets fried, the rest can carry on. Resilience kind like digital duct tape. As more cars and stoplights chat via V2I, there’s a big, flashing red light called “cybersecurity.” Hackers already mess around with highway signs for laughs imagine what they could do here. Introducing crazy stuff like block chain for validation and smart algorithms to sniff out weird activity isn’t just overkill; it’s starting to look like basic common sense. One last thing can we please get some standards. Right now, every study uses its own weird yardstick. If we are ever a know what’s the “best” 77 | P a g e DOI: 10.5281/zenodo.17359016 algorithm, we need common datasets and rules. Also, wouldn’t hurt to mix in a healthy dose, human reality like modelling how actual people drive/walk, not just the idealized robots of academic papers. This isn’t just tech for tech’s sake it’s shaping whether cities will be places you want to live in, not just drive through. Summary To improve city traffic by using advanced computer systems to manage traffic signals. The main goal is to make traffic flow faster, reduce accidents, and decrease pollution. The paper explains that traditional traffic lights are not effective because they can't adapt to changing traffic conditions. Instead, it proposes a new approach where systems learn from traffic patterns and car communication to make real-time decisions. This is achieved using different learning methods and setups. The document highlights that current research often focuses on only one of the three goals (efficiency, safety, or environmental impact), but a more balanced approach is needed for realworld cities. It also points out that most of this research has been done in simulations and has not been tested on a large scale in actual cities. The paper proposes a framework for future implementation, including a multiobjective approach that considers electric vehicles and real-world validation. Conclusion After through all the new stuff on DRL-driven adaptive traffic signal controls, got to say it's wild how far we’ve come. Smarter lights, greener cities, less cussing on your morning commute. If you stack DRL systems up against the old-school timers or even that fancy “actuated” setups, it’s not even close. DRL just rolls with the punches, making calls in real time, like it knows when you’re late for work. Less stop-and-go means folks get places faster, fewer fender-benders, and hey, we’re not choking out the planet quite as much (science says). Plus, once you toss more Connected and Autonomous Vehicles (CAVs) into the traffic soup, these brainy systems really start cooking. The more our cars talk to the lights and each other the smoother everything flows. It's almost like well, just modern engineering flexing. Not all sunshine and green lights here. There are still a few potholes in the road. First up, so many papers chase just one goal: "Let’s kill emissions!" or "Let’s make it super safe!" or "Efficiency at all costs!" In the real world, you can’t just pick one lives messy. Juggling safety, emissions, and keeping things speedy That’s the real trick, and we’re not quite there yet. In a simulation tends to fall apart when you dump it on an actual city with thousands of intersections and drivers who treat traffic laws as polite suggestions. Also reward design. If you try to make the system only care about CO₂ numbers, you get these weird, twitchy results. Instead, we end up telling the AI, “Just keep the lines short, stop slamming on the brakes, you’ll figure out the rest. “Simulations SUMO, VISSIM all nice playgrounds but they’re not real life. No weather, no surprise construction, no “that guy” blocking the intersection because he had to beat the yellow. We need way more street tests, not just digital ones. Right now, real world pilot studies are rarer than a green light at rush hour, and without these systems are stuck in science fair mode. And don’t even get me started on the sticky stuff: privacy, ethical headaches, fairness, and who's liable when the whole system glitches out. That’s a can of worms the tech folks and city planners really need to crack open. DRL and CAVs together could honestly make urban traffic something we complain about way less, if we do it right. But only if 78 | P a g e DOI: 10.5281/zenodo.17359016 folks build systems that juggle all the real-world stuff scale up smart, test them on actual streets, and balance out the big three safety, efficiency, and saving the planet. Could be the start of traffic systems that don’t totally suck and maybe even help us live in legit smart cities. References [1]. Wang, Z., Xu, L. and Ma, J. (2023). Carbon Dioxide Emission Reduction-Oriented Optimal Control of Traffic Signals in Mixed Traffic Flow Based on Deep Reinforcement Learning. Sustainability, 15(16564). [2]. Kang, Z., An, L., Yang, X., and Lai, J. (2024). A Carbon Benefits-Based Signal Control Method in a Connected Environment. Applied Sciences, 14(7638). [3]. Karbasi, A. H., Yang, H., and Razavi, S. (2024). Exploring the Impact of Traffic Signal Control and Connected and Automated Vehicles on Intersections Safety: A Deep Reinforcement Learning Approach. McMaster University. [4]. Mir bakhsh, S., and Azizi, M. (2024). Adaptive Traffic Signal’s Safety and Efficiency Improvement by Multi-Objective Deep Reinforcement Learning Approach. IJIRME, 3(7), 1245–1257. [5]. Guo, J., Cheng, L., and Wang, S. (2023). COTV: Cooperative Control for Traffic Light Signals and Connected Autonomous Vehicles using Deep Reinforcement Learning. IEEE Transactions (preprint). [6]. Tang, C. R., Hsieh, J. W., and Teng, S. Y. (2023). Cooperative Multi-Objective Reinforcement Learning for Traffic Signal Control and Carbon Emission Reduction. Neur IPS 2023. [7]. Schumacher, M. E. H., Adriano, C. M., and Giese, H. (2023). Challenges in Reward Design for Reinforcement Learning-based Traffic Signal Control: An Investigation using a CO₂ Emission Objective. SUMO Conference Proceedings, 4(222). [8]. Mascia, M., Hu, S., Han, K., North, R., Van Poppel, M., Theunis, J., Beckx, C., and Litzenberger M. (2016). Impact of Traffic Management on Black Carbon Emissions: A Microsimulation Study. Networks and Spatial Economics, 17(269–291). [9]. Talluri, K. K., Stang, C., and Weidl, G. (2025). Green Wave as an Integral Part for the Optimization of Traffic Efficiency and Safety: A Survey. IEEE Intelligent Vehicles Symposium (IV) 2025. [10]. Kim, J., Jung, S., Kim, K., and Lee, S. (2019). The Real-Time Traffic Signal Control System for the Minimum Emission using Reinforcement Learning in V2X Environment. Chemical Engineering Transactions, 72(91–96).