Post on 31-Oct-2021
Two-Way Transit Signal Priority Algorithm for
Optimizing Transit Reliability and Speed A Deep Reinforcement Learning Approach
Wen Xun Hu, Chuhan Chen, Hirotaka Ishihara, Islam Taha, Amer Shalaby and Baher Abdulhai
Motivation
The transit
reliability and
speed are
performance
indicators
Transit
services are
vulnerable to
variability and
delays
No strategies can
adaptively improve
headway regularity
and reduce signal
delays
simultaneously
The conventional
TSP only aims at
reducing delays
Multiple TSP
requests from
different
approaches are
commonly
handled by
FCFS logic
Model of the
dynamics of
transportation
environment
Objectives
▪ Dual-objective TSP
– Adaptively optimize reliability (i.e., headway regularity) and
reduce signal delays simultaneously
▪ Coordination of opposite directions
– Develop an algorithm to coordinate TSP in opposite directions
of the same intersection based on real-time bus performance
Objectives
▪ High speed, poor reliability
– Pros
• Reduced in-vehicle travel time
– Cons
• Long waiting time
• Uneven bus loads (e.g., crowding)
• Operating costs
• etc.
Source: https://www.thestar.com/news/gta/2020/11/27/the-ttc-wants-to-test-
platoons-of-driverless-buses.html
https://www.thestar.com/opinion/star-columnists/2017/12/15/are-we-sure-ttc-fare-
evasion-happens-as-often-as-we-think-it-does.html
Objectives
▪ Reliability vs. Speed
– One directionScenario A
A bus arrives at the intersection with a headway (hin) > scheduled headway (hs).
• To improve speed and reduce headway deviation, the TSP system should expedite this bus.
Scenario B
hin < hs:• To improve speed or reduce signal delays, the TSP system should expedite this bus.
• To reduce headway deviation, the TSP system should hold back (i.e., slow down) this bus.
– Opposite directions• More conflicting scenarios
Model FormulationOne-Way DRL Agent
POZ
Check-out
Loop detector
Check-in
Loop
detector
• Model-free deep reinforcement learning
• Efficient for large state space
Model FormulationOne-Way DRL Agent
• Target travel time to maintain scheduled
headway at check-out
• Number of vehicles in the POZ
• Time to the end of 1st available green
reward =
𝑤ℎ ℎ𝑒𝑎𝑑𝑤𝑎𝑦 𝑖𝑚𝑝𝑟𝑜𝑣𝑒𝑚𝑒𝑛𝑡 − 𝑤𝑡(𝑡𝑖𝑚𝑒 𝑖𝑛 𝑃𝑂𝑍)
Transportation network
simulated in Aimsun
TSP control system
Green phase:
-20, -15, -10, -5, 0, +5,
+10, +15, +20 sec
←RL image:
https://shield.ai/content/2018/11/20/shield-ai-
fundamentals-on-reinforcement-learning
Source: Pawel Czerwinski,
unsplash.com
Signal cycle image:
https://winnipeg.ca/publicworks/transportati
on/trafficsignals/trafficsignaloperation.stm
Model FormulationCoordination Algorithm
RL agent X
Coordination
algorithm
𝑄 𝑠𝑡 , 𝑎; 𝜃𝑡
𝑄 𝑠𝑢 , 𝑎; 𝜃𝑢
𝐀𝐜𝐭𝐢𝐨𝐧: 𝒂𝒕′
= 𝑎𝑟𝑔𝑚𝑎𝑥𝑎
𝑤𝑥𝑄 𝑠𝑡 , 𝑎; 𝜃𝑡 + 𝑤𝑦𝑄 𝑠𝑢, 𝑎; 𝜃𝑢RL agent Y
Simulation case study
▪ Aimsun Next Microsimulation
▪ Intersection in Toronto
▪ Bus line: 36 Finch West
Finch Ave West at Kipling Ave
Results
▪ Base scenarios
– No TSP
– TSP in field (Toronto TSP)
– DRL agents + FCFS logic (FCFS TSP)
▪ Proposed scenario
– DRL agents + coordination algorithm (D2 TSP)
Source: Pawel Czerwinski,
unsplash.com
Source: Andre Furtado, unsplash.com
Results
Coefficient of Variation of Headway
0.461
0.439
0.4170.408
0.502 0.505
0.482
0.456
0.40
0.42
0.44
0.46
0.48
0.50
No TSP Toronto FCFS D2 TSP
EB WB
Results
% of Extreme Headways
3.4%
18.3%
22.1%
8.3%
3.8%
16.8%
20.8%
6.7%
1.7%
13.9%
16.9%
7.3%
1.2%
13.1%
15.4%
5.9%
0.0%
5.0%
10.0%
15.0%
20.0%
25.0%
h < 0.25 hs h < 0.5 hs h > 1.5 hs h > 1.75 hs
No TSP Toronto FCFS D2 TSP
Results
Travel Time90.7
66.4
77.9
73.1
91.3
60.257.7
62.5
55.0
60.0
65.0
70.0
75.0
80.0
85.0
90.0
95.0
No TSP Toronto FCFS D2 TSP
EB WB
9 s travel time ↑,
23 s stdev of h dev↓ = 10 s waiting time↓
Conclusions
The proposed Two-Way TSP (D2 TSP)
▪ Generates the best headway performance in both directions
– Effective in reducing
headway variability and
% of extreme headways
▪ Brings noticeable reduction in travel time compared with “No TSP”
Source: Kayla Speid, unsplash.com
Future Work
▪ Coordinated TSP systems to enhance the benefit on transit reliability and speed at the route level
▪ Use connected vehicle technologies for detection and communication
▪ Integrate TSP design with other route strategies
Source: Nong Vang, unsplash.com
THANK YOU!
wenxun.hu@mail.utoronto.ca
16
Source: chuttersnap, unsplash.com