Adaptive Signal Control: Cyclic, or not?
This post has already been read 1025 times!
At LinkedIn Group Adaptive Traffic Signal Control Systems, Dr. Andrei Reztsov, an applied mathematician posted a link for his new paper
Self-Organizing Traffic Lights as an Upper Bound Estimate
Andrei Reztsov, Complex Systems 24(2).
Self-organizing traffic lights (SOTL) are considered a promising instrument for the development of more adaptive traffic systems. In this paper we explain why some well-promoted results obtained with the use of SOTL should be scrutinized and carefully reviewed. Current computational research projects based on SOTL should be reviewed too.
For those interested, Dr. Reztsov’s papers can be downloaded from SSRN.
I applaud this paper for its insights and for its logical and rigorous treatment from an mathematician to point out the methodological fallacies of SOTL or those type of second-by-second genre of real time signal control. I concur with the opinion of the paper. As far as I know, this appears the first to discuss the following aspects of adaptive signal control:
- cyclic calculation/decision,
- sub-cyclic calculation/decision, and
- sec-by-sec decision.
Self-Organizing Traffic Light (SOTL), or sometimes “Second-By-Second” control refers to acyclic operations by internally manipulating phase force-off and phase hold commands so the signals do not operate on a cyclic basis. It is “second-by-second”, in the sense the system monitors phases status and detectors inputs continuously, and makes decisions to terminate or extend phases on a second-by-second or near second-by-second basis. These are currently not within standard NTCIP C2F operations.
A centralized second-by-second control typically would result in high communication overhead due to the onerous second-by-second status monitoring at each individual phase and/or detector level. Also second-by-second control cannot be easily integrated with current NTCIP Center-2-Field framework, because NTCIP is UDP based type of communication that cannot guarantee the receipt of data packets, thus not meeting the reliability requirement for second-by-second control. In practice it is typically implemented in a distributed manner at local controllers, and requires certain middle-tier mechanism (e.g., a field master, or a customized firmware embedded in the same cabinet, or an add-on module communicating with a standard SDLC port) to coordinate with adjacent intersections and cache the second-by-second commands without overwhelming the central system, if any.
Second-by-second type of control is flexible, and effective when traffic demand is low to the extend of random arrivals. It is most effective when the network starts with empty streets – that is exactly where this type of control’s niche resides to best reshuffle the time resource – to accommodate the predicted arrival of platoons on a preferred route. It works well, only and only when there is time resource to exercise such “rescheduling”.
And that is the problem.
The traffic signal optimization problem is fundamentally simple – it boils down to allocate either limited time resource (for oncoming vehicles), or limited space resource (for queuing vehicles), of at-grade intersections with competing traffic streams.
“At-grade intersections” means the system has to deal with competing traffic streams in a 2-D plane. Both time and space resources are limited for at-grade intersections. Time resource is limited, because in practice any at-grade intersection’s capacity will never exceed 1800 vphpl; space resource is limited, because it is constrained by available storage space.
The signal for each phase applies to the group of drivers, not an individual driver on an individual stop-release basis. When traffic is light, the signals running second-by-second can favor the predicted on-coming traffic on a preferred route, thus reducing perceived maximum waiting time for individual drivers, and improving individual driver satisfaction.
The challenge really comes when traffic becomes heavy and over-saturated. In that case, the available capacity of an intersection is not able to serve the demand, and the flexibility to shuffle the time resource for vehicle platoons is gone.
It is logical to believe that when traffic is light to the extent of pure random Poisson arrivals, SOTL behaves more like an enhanced actuated type of control; when traffic increases, SOTL would converge to cyclic no matter how the logic tries to reshuffle the time for individual vehicles. The chain of reasoning is:
When traffic is light, the value of sec-by-sec may help individual driver satisfaction due to its flexibility to freely terminates a phase, and due to the reduced MAX perceived delay time, but not primarily average delay. However, when demands increase, the benefits quickly diminish and average delay increases. At certain range of traffic flow regime, sec-by-sec control probably reduces capacity because of the increased lost-time from phase switches. Remember, phase switches have a cost of lost time which is non-trivia when traffic is not light. Therefore, with traffic increasing, sec-by-sec will quickly converge to cyclic losing its point.
This explains my particular favor of a cyclic system and I found this paper is very interesting.
Convergence to cyclic and the effect of sec-by-sec of reducing MAX perceived delay instead of AVG delay renders such operation moot in practice, because:
- When traffic is light as random arrivals, there is really not much need or systematic benefits of running adaptive. This is because a fully actuated setting can well handle light traffic conditions. Favoring a route when traffic is light may well promote speeding and incur other implications.
This also explains why sec by sec control by itself wouldn’t be working well for congested or heavy traffic, at least not as well as some of the literature reported. It appears to me many before-and-after improvements were built on the base case of comparisons being poorly-tuned fixed time plans, or not well-configured semi-coordinated operations.
In summary, using comments posted in the same LinkedIn Group, from Mr. Kevin Fehon:
It is logical that SOTL or similar systems must converge to a cyclical state with heavy demand and practical constraints (such as maximum wait time, driver expectations about fairness in distribution of delays, etc.). This is built-in in their operations, and is supported by observation of existing non-cycle based systems in the field. It appears that this convergence also cannot be as optimized as the operation of a system whose optimization assumes cyclical operation, especially when vehicle-actuated flexibility (phase re-service, phase sequence changes under different traffic conditions, etc.) is built into the system.
This post has already been read 1025 times!