基於深度Q網路的交通號誌燈週期配置

Traffic Light Cycle Configuration Based on Deep Q Network

廖翌翔(YI-XIANG LIAO)

指導教授 ： 朱鴻棋

朝陽科技大學/資訊學院/資訊與通訊系/碩士(2020年)

摘要

近年來，人口密度迅速的增加，而人口密度高的國家則飽受交通壅塞的困擾。嚴重的交通壅塞會導致許多問題的產生，例如車輛廢氣排放、車輛旅行時間增加、能源消耗和車輛事故的產生。而導致交通壅塞問題的因素很多，包含道路的通行能力不佳、車輛密度過高、不佳的城市交通規劃或無法因應實際車流的交通號誌燈設置。而交通號誌燈的週期配置為重要的因素之一。交通號誌燈的週期如果能因應車流動態調整，能夠有效地降低交通壅塞程度。因此我們提出了一種基於Q-learning的改進機制來優化交通號誌燈週期配置，以較低運算時間和接近最佳的車輛平均延遲時間，並因應實際車流而調整交通號誌燈的週期配置。實驗結果表明，所提出的方法處理步驟優於窮舉搜索法11.76倍，而且車輛平均延遲時間僅比窮舉搜索法略低5.4%。而交通壅塞問題會影響鄰近的路口，本文使用Deep Q network方法動態調整鄰近路口的交通號誌燈週期配置。經實驗結果顯示，所使用的方法的總車輛延遲時間在最佳情況下優於固定式週期配置12.86%。本文所提出的方法能夠通過控制交通號誌燈有效舒緩交通壅塞問題。

關鍵字

交通壅塞 ； 交通號誌燈週期配置 ； 機器學習 ； 強化學習 ； Q學習 ； 深度Q網路

並列摘要

In recent years, population density has increased rapidly and countries with high population density have been plagued by traffic congestion. Traffic congestion can cause many problems, such as vehicle exhaust emissions, increased vehicle travel time, energy consumption, and vehicle accidents. There are many factors leading to traffic congestion, including poor road capacity, high vehicle density, poor urban traffic planning, or insensitive traffic light settings. The cycle configuration of traffic lights is one of important factors. If the cycle of traffic lights can be dynamically adjusted in response to the traffic flow, it can effectively reduce traffic congestion. Therefore, we propose an improved mechanism based on Q-learning to optimize the traffic light cycle configuration, with a lower calculation time and close to the best vehicle average delay time, and adjust the traffic light cycle configuration according to the actual traffic. Experimental results show that the processing steps of the proposed method are 11.76 times better than the exhaustive search method, and the average vehicle delay time is only slightly lower than the exhaustive search method by 5.4%. The problem of traffic congestion will affect adjacent intersections. In this paper, the Deep Q network method is used to dynamically adjust the traffic light cycle configuration of adjacent intersections. The experimental results show that the total vehicle delay time of the proposed method outperforms the fixed period configuration by 12.86% in the best case. The proposed method can effectively relieve traffic congestion by controlling traffic lights.

並列關鍵字

Traffic Congestion ； Traffic Light Cycle Configuration ； Machine Learning ； Reinforcement Learning ； Q-Learning ； Deep Q Network

參考文獻

參考文獻

[1] “中華民國內政部戶政司全球資訊網.” [Online] https://www.ris.gov.tw/app/portal/346, (Accessed on 22 June 2020).

[2] “智慧運輸系統發展建設計畫.” [Online] https://www.ey.gov.tw/Goals/4BDAC992C15E36DA, (Accessed on 22 June 2020).

[3] “臺中市恢復BRT恐導致交通惡化情形專案報告.” [Online] https://www.rdec.taichung.gov.tw/12047/12142/12145/1054952/1055303, (Accessed on 22 June 2020).

[4] “臺灣混合型車流是示意圖.” [Online] https://news.ltn.com.tw/news/society/breakingnews/1407099, (Accessed on 22 June 2020).