Academia.eduAcademia.edu

Figure 12 – uploaded by Teerapat Sanguakotchakorn

See full PDFdownloadDownload figure
Figure 12 shows the average waiting time delay with and without deep reinforcement learning approach over 15 runs versus traffic density. It is apparent that the average waiting time delay has similar trend as the average travelling time delay in figure 10. The comparison results illustrate explicitly the effectiveness of the algorithms when traffic density is more than 2,500 vehicles/km?, which MLP algorithm is still the most effective one.

Figure 12 shows the average waiting time delay with and without deep reinforcement learning approach over 15 runs versus traffic density. It is apparent that the average waiting time delay has similar trend as the average travelling time delay in figure 10. The comparison results illustrate explicitly the effectiveness of the algorithms when traffic density is more than 2,500 vehicles/km?, which MLP algorithm is still the most effective one.

Related Figures (12)

International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.4, July 2021 Figure 1. Flowchart Diagram of our Proposed Algorithm
International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.4, July 2021 Figure 1. Flowchart Diagram of our Proposed Algorithm
For each simulation, SUMO generates the number of outputs in XML format, including each vehicle’s position in each simulation step, inductive loops, information of all roads and lanes and travelling trip information for each vehicle. Not only the measurement of conventional traffic, but SUMO also has additional options for fuel consumption, together with pollution or noise emission model. Figure 2. Traffic grid network
For each simulation, SUMO generates the number of outputs in XML format, including each vehicle’s position in each simulation step, inductive loops, information of all roads and lanes and travelling trip information for each vehicle. Not only the measurement of conventional traffic, but SUMO also has additional options for fuel consumption, together with pollution or noise emission model. Figure 2. Traffic grid network
Figure 3. Real-world road network
Figure 3. Real-world road network
Figure 4. Flowchart of traffic congestion prediction concept However, in the first program running, it might not be able to choose the best path which results in the smallest amount of travelling time delay and waiting time delay. Therefore, the program will be run more than once. As the data of action taken is memorized in the model, the higher number of programs running will increase the prediction efficiency of the congested path, which results in the decrement of the travelling time delay and waiting time delay. This process is kept repeating until the time delay does not decrease, which means it reaches the minimum value.
Figure 4. Flowchart of traffic congestion prediction concept However, in the first program running, it might not be able to choose the best path which results in the smallest amount of travelling time delay and waiting time delay. Therefore, the program will be run more than once. As the data of action taken is memorized in the model, the higher number of programs running will increase the prediction efficiency of the congested path, which results in the decrement of the travelling time delay and waiting time delay. This process is kept repeating until the time delay does not decrease, which means it reaches the minimum value.
International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.4, July 2021
International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.4, July 2021
Figure 6. Flowchart diagram of simulation procedure
Figure 6. Flowchart diagram of simulation procedure
Figure 7. Deep reinforcement learning All deep reinforcement learning models use the similar concepts of agents, environments, states, actions and rewards. In figure 7, the deep reinforcement learning process is illustrated. The element called environment is responsible to convert an action taken in the particular state into a new state and reward. After that, the result values will be altered again by an agent which transforms them into the next action. An agent attempts to take actions in the environment, providing the maximum rewards.
Figure 7. Deep reinforcement learning All deep reinforcement learning models use the similar concepts of agents, environments, states, actions and rewards. In figure 7, the deep reinforcement learning process is illustrated. The element called environment is responsible to convert an action taken in the particular state into a new state and reward. After that, the result values will be altered again by an agent which transforms them into the next action. An agent attempts to take actions in the environment, providing the maximum rewards.
The complexity of our proposed algorithms is also analyzed using Big O notation, which is the method using to describe the complexity and execution time required by an algorithm [9]. In this deep reinforcement learning model, the maximum time that the algorithms can perform is in cubic time written as O(n*), where n is the size of inputs. That means the running time is proportional to the cube of the input size. In figure 9, the Big-O complexity of this algorithm in the operation unit (x10°) is illustrated. The input size is the number of vehicles in the traffic simulation.
The complexity of our proposed algorithms is also analyzed using Big O notation, which is the method using to describe the complexity and execution time required by an algorithm [9]. In this deep reinforcement learning model, the maximum time that the algorithms can perform is in cubic time written as O(n*), where n is the size of inputs. That means the running time is proportional to the cube of the input size. In figure 9, the Big-O complexity of this algorithm in the operation unit (x10°) is illustrated. The input size is the number of vehicles in the traffic simulation.
Table 2. Simulation Parameters 4.7. Performance Evaluation Metrics International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.4, July 2021 The performance of our proposed algorithm is evaluated by the following metrics:
Table 2. Simulation Parameters 4.7. Performance Evaluation Metrics International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.4, July 2021 The performance of our proposed algorithm is evaluated by the following metrics:
between the results with three deep learning and those results without deep learning algorithms are obvious. The deep reinforcement learning model apparently becomes more effective when the raffic density is around 3,500 — 4,500 vehicles/km?. If the traffic density continues to increase until it reaches 5,000 vehicles/km’, the rerouting seems to be less possible, since there are not enough available roads to move. Based on the results shown in figure 10, it can be concluded that he most effective deep learning algorithm for traffic congestion prediction model is MLP, followed by CNN, while the least effective one is LSTM. In Figure 11, it shows the reduction percentage in average travelling time delay after using three different deep reinforcement learning algorithms. It is obvious that MLP provides the highest percentage of average travelling time reduction among all three algorithms.
between the results with three deep learning and those results without deep learning algorithms are obvious. The deep reinforcement learning model apparently becomes more effective when the raffic density is around 3,500 — 4,500 vehicles/km?. If the traffic density continues to increase until it reaches 5,000 vehicles/km’, the rerouting seems to be less possible, since there are not enough available roads to move. Based on the results shown in figure 10, it can be concluded that he most effective deep learning algorithm for traffic congestion prediction model is MLP, followed by CNN, while the least effective one is LSTM. In Figure 11, it shows the reduction percentage in average travelling time delay after using three different deep reinforcement learning algorithms. It is obvious that MLP provides the highest percentage of average travelling time reduction among all three algorithms.
Figure 12. Comparison of the Average Waiting Time Delay between the algorithms with and without deep reinforcement learning approach over 15 runs
Figure 12. Comparison of the Average Waiting Time Delay between the algorithms with and without deep reinforcement learning approach over 15 runs
Related topics:
Computer ScienceDistributed ComputingReinforcement LearningSUMO vectorVehicular ad hoc networkWireless Ad Hoc Network
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved