TY - GEN
T1 - From Local to Global
T2 - 2nd IEEE International Conference on Software Engineering and Artificial Intelligence, SEAI 2022
AU - Zheng, Nianzhao
AU - Li, Jialong
AU - Mao, Zhenyu
AU - Tei, Kenji
N1 - Funding Information:
ACKNOWLEDGMENT The research was partially supported by JSPS KAKENHI and JSPS Research Fellowships for Young Scientists. The authors would like to thank Mr. Kun Liu for his comments on the early stage of this study.
Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - Traffic signal control (TSC) is one of the effective ways to mitigate traffic congestion, a growing problem causing significant economic loss to urban areas. In recent years, there is an increasing interest in using reinforcement learning (RL) in TSC since RL shows great potential in optimizing control policy for complex real-world traffic conditions. However, one concerning problem is the huge computing time/resources required to train the control policy for large-scale TSC. To address this problem, this paper proposes a method that utilizes a curriculum to help speed up the training process. Control policies of different single-intersection maps are learned first and then referenced to a large-scale map consisting of a certain number of different intersections. The preliminary evaluation demonstrates that our method achieves a jump-start compared to that of learning the target task from scratch.
AB - Traffic signal control (TSC) is one of the effective ways to mitigate traffic congestion, a growing problem causing significant economic loss to urban areas. In recent years, there is an increasing interest in using reinforcement learning (RL) in TSC since RL shows great potential in optimizing control policy for complex real-world traffic conditions. However, one concerning problem is the huge computing time/resources required to train the control policy for large-scale TSC. To address this problem, this paper proposes a method that utilizes a curriculum to help speed up the training process. Control policies of different single-intersection maps are learned first and then referenced to a large-scale map consisting of a certain number of different intersections. The preliminary evaluation demonstrates that our method achieves a jump-start compared to that of learning the target task from scratch.
KW - curriculum learning
KW - reinforcement learning
KW - traffic signal control
UR - http://www.scopus.com/inward/record.url?scp=85136333754&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85136333754&partnerID=8YFLogxK
U2 - 10.1109/SEAI55746.2022.9832372
DO - 10.1109/SEAI55746.2022.9832372
M3 - Conference contribution
AN - SCOPUS:85136333754
T3 - 2022 2nd IEEE International Conference on Software Engineering and Artificial Intelligence, SEAI 2022
SP - 253
EP - 258
BT - 2022 2nd IEEE International Conference on Software Engineering and Artificial Intelligence, SEAI 2022
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 10 June 2022 through 12 June 2022
ER -
