Intelligent traffic control for autonomous vehicle systems based on machine learning Note
Lee, S., Kim, Y., Kahng, H., Lee, S. K., Chung, S., Cheong, T., … & Kim, S. B. (2020). Intelligent traffic control for autonomous vehicle systems based on machine learning. Expert Systems with Applications, 144, 113074.
Introduction
Real-world traffic problem in a large scale plant
Traffic control is a significant challenge with Autonomous Vehicle Systems (AVS) because all vehicles must be monitored and controlled in real-time to cope with uncertainties such as vehicle overcrowding on the railway exceeding transportation.
Existing traffic control systems are insufficient to prevent heavy congestion that impedes production. Thus, developing a predicitve control system is an important issue for semicoducotor manufacturing.
Literature Review
Traffic Problems
(1) Optimization of the layout design of AVS
The fundmental solution of traffic problem, however, once the layouts has been put in place, it is difficult to revise.
(2) Operational issues
Including dispatching rules, idle vehicle policies, vehicle routing planning, all the studies rely on assumption, vehicle operational issues under steady-state production condition, cannot reflect the dynamic production conditions.
(3) Optimizing the scheduling and routing of vehicles
Including heuristic optimization methods, metaheuristic algorithms, simulation studies, and the methods foused on detection of deadlock and livelock problem caused by interference. But these studies only considered small problems involving few vehicles.
Congestion
(1) Uncertainty inherent in designing vehicle routes
Including node-based updating mehod and arc-based upadting algorithm. Both approaches are not impraticle because of excessive load for system resources when recalculating short-cycle route and the effectiveness of real-time rerouting is trivial.
(2) Heuristic approach
Focused on routing with real-time updates of the panalty parameters, however, these must be analyzed and devised by human experts.
(3) Machine learning approach
To optimize wafer transfer and leverage machine utilization, several approaches used reinforcement learning. However, these approaches have had limited opportunities for application to large-scale problems and human intervention is still essential to build well-trained model.
Predictive Traffic Control System
Data description
Used a simplified version of a real-world railway network to produce a graph-based representation containing nodes and arcs.
Nodes are loading and unloading ports that are connected to production facilities and rail intersection points. Arcs are unidirectional rails.
To reduce the complexity of the network connectivity and achieve data consistency, converted the network to an intersection point network.
Then calculated traffic indices for each section, as identified by the branching and merging points on the network structure.
The driving logs of vehicles, this study obtained three days of real data from a fab to estimate traffic condition. The columns included timestamps, driving states, locations, speed, assigned transfer job, and designate vehicle routes.
Predicitve model
Used a machine learning-based predictive model to predict two traffic indexes. Compared four widely-used models: Random Forest(RF), Support Machine Vector(SVM), Gradient Boosting Machine(GBM) and Deep Neural Network(DNN).
Traffic control model
Conducting two different predicitve model for traffic volume and speed.
The predictive model for traffic volume was used to verify whether traffic volume fluctuations caused congestion, and the predictive model for speed was used to identify the level of congestion.
Determined the degree of congestion level: free, busy and congested.
(1) Free: a section has low traffic volumes with large variations in average speed.
(2) Busy: The speed deceleration starts with increasing vehicle volume.
(3) Congested: The speed drastically decreases because of extreme traffic volume.
Using change-point detection(CPD) to detecting the state change in the speed. A change point is determined as the moment when sequential hypothesis tests detect the probability distribution of data changes.
Using the Brown-Forsythe test(F-test) which is useful in indicating a statistically significant difference in variances across different groups, uses the F statistic to compare absolute deviations from the median between different groups.
The proposed predicitve traffic control system contains traffic control and predicitve models. The proposed system performs prediction and control once every five minutes. Using Dijkstra’s algorithm based on shortest-path problem to determine the vehicle’s route. For message communication, this study used commercial message-oriented middleware solution, Tibco Rendezvous, for scalability and robustness (TIBCO Rendezvous TM Concepts, 2002), to broadcast information about the size of traffic from the existing routing system to the traffic control model.
The routing configuration parameters correspond to the penalty costs of the sections, which were used as the distance criteria for route calculation. If a penalty cost increases, the section is more likely to be excluded from the route selection. The updated penalties are only applied to newly-assigned vehicles.
Experiments
Traffic prediction results
Training data sets: first and second-day data.
Predicitng data sets: third-day traffic indices.
Predictive performance: Mean Absolute Percentage Accuracy(MAPA)
Simulation
Simulation contains multiple vehicle emulators and system communicaton, transfer job assignment and autonomous driving. Conducting 10 replications with differnet random seeds and measured the average of the performance measures, collected once per hour in each replication.
Results
Three evaluation criteria: delivery time, transfer time and queued time.
(1) Delivery time = queued time before assigning a wafer lot to vehicles + transfer time
(2) Transfer time = elapsed time between the unloading of the wafer lot at the departure machine when it reaches its destination
(3) Queued time = the time from the delivery request to the assignment of a vehicle
Conducting Kolmogorov-Smirnov(KS) tests to statistically assess whether the rank of the population mean differed between the existing Dijkstra-based shoretest path planning system and proposed predicitve traffic control system.
The p-values for more significant bottleneck sections were less than 0.001, indicating that the performance difference between the existing algorithm and the proposed algorithm is statistically significant.
Conclusion
(1) Developing a machine learning-based predictive traffic control method for AVSs in manufacturing, especially for high-workload situation.
(2) The proposed approach demonstrated improvement in steady-state performance in delivery and transfer times, and elimination of congestion in significant bottleneck sections.
(3) Planning to apply the predictive traffic control system to medium- and large-sized 300 mm wafer facilities in a Korean semiconductor manufacturing company(Samsung).
Limitaion
(1) Assumed stationary traffic conditions for AVSs, however, traffic pattern changes are inevitable in real-world problems.
(2) This study built independent predictive models for each bottleneck section. However, significant bottleneck sections are expected to be correlated, because the sections are connected directly or indirectly in railway networks.
(3) The proposed method is only designed to alleviate short-term congestion, on the order of five minutes.
(4) The approach method was ineffective in the sections with less congestion. To overcome this, we need to facilitate more precise traffic control according to the expected level of congestion.
