Y.W. Almulla, G.M. Brandt
Oregon State University,
Keywords: self-driving, highway safety, human drivers, intelligent-agent modelling, traffic modelling
Summary:The field of intelligent-agent modeling is becoming increasingly relevant as time goes on, especially due to recent advances in machine learning and artificial intelligence. This field is of particular interest to industry, as the market for self-driving cars (SDCs) and other intelligent devices is rapidly growing. Remarkably, little work has been publicly published in the field of SDC computational modeling; in 2016, Gora and Rub published a work concerning modern methods for successfully modelling systems with SDCs, but go on to highlight that none have been completed yet. This work extends upon their concepts and presents a model for simulating traffic conditions on roads with mixed amounts of SDCs and human-driven cars (HDCs). SDCs have faster, more accurate reactions to their environments than HDCs, and also have information-sharing capabilities that set them apart for their potential to increase the safety of national highway systems. Our model uses complex decision trees to account for these differences in behavior, guiding the HDCs based on local conditions alone, but both local and global conditions for the SDCs. The difference in behavior for SDCs and HDCs is made explicit in the logical conditions required for certain actions, such as accelerating and switching lanes. We motivate the use of our model and justify its accuracy by showing that it models the traffic density conditions of road conditions when there are 0% SDCs. Note that adding 30% SDCs means 30% of the vehicles in the I-5 simulation are SDCs, and 70% are HDCs. Interstate I-5 is chosen as the model highway for this work, but the generality of our computational model allows it to be employed for any highway system. We present preliminary results for simulating the addition of 30%, 60%, 90%, and 100% SDCs to Interstate Highway I-5. We discuss the flexibility of our model in being machine learned to optimize for any desired traffic property, and the potential strengths and weaknesses of road systems which are dependent on a cyber-infrastructure. Currently we are applying machine learning methods to optimize the behavior of the SDCs for reducing time of travel and average crash count. An SDC highway partitioning scheme (SDC-HPS) is presented as a method to garner many of the benefits of having a highway system dominated by SDCs. This concept is argued to be optimal because it could begin within one year at an estimated cost of $500,000-$1,000,000, whereas the alternative method to acquire similar levels of highway safety is to wait until the majority of citizens own SDCs themselves. This could take decades, as SDC technology is not yet advanced enough and citizens will not feel the economic incentive to switch to SDCs for a similar amount of time. In contrast, the SDC-HPS does not require very advanced self-driving technology, only the ability for vehicles to be controlled from long-distance.