Traffic Lights Strategy Adaptation

Veiw figure View Figure

Simulation

In this phase we have to conduct a necessary amount of experiments when a behaviour of the system is observed. Characteristic patterns of the system behaviour are searched in its output indicators. These patterns are saved and described. Particular patterns are determined. These patterns are split into several categories, each of which has similar properties. A training set is created by this procedure. For every category, it is necessary to define what will happen if any pattern from any category emerges in the system’s behaviour.

Behaviour Pattern Recognition

In this phase the neural network is adapted by the prepared training set. This task makes use of our own developed tool called EDU Sandbox (https://sourceforge.net/projects/esbox/). Then the neural network is able to generalise the learned training set. After that, when the simulation is running, the neural network can recognize similar behaviour pattern it already knows and returns that information back to the simulation software.

Adaptation of Parameters

Based on the recognised patterns in the system’s behaviour it is possible to adapt action parameters’ values according to the prepared expectations. Adaptation logic has to be programmed in the simulation software.

The flowchart of the methodology (Figure 1) represents a more detailed step-by-step procedure with each above-mentioned phase.

3. Experiment

For our experiment we use a simulation prepared by Carlos Gershenson ^[7] which is available at https://turing.iimas.unam.mx. We have modified this simulation to able to connect it to our pattern recognition utility. Adaptation logic was added as well. The model simulates city traffic where traffic lights try to “self-organise” to improve the traffic flow in the city. It consists of an abstract traffic grid with intersections between cyclic single-lane arteries of two types: vertical or horizontal. Cars only flow in a straight line, either eastbound or southbound. Each crossroad has traffic lights that allow traffic flow in only one of the intersecting arteries with a green light.

The experiment used the following tools: NetLogo simulation software ^[8], MySQL database, EDU Sandbox. Communication between these tools is following (Figure 2).

Figure 2. Communication between NetLogo and EDU Sandbox

Download as

Veiw figure View Figure

The simulation runs in the NetLogo simulation software. The simulation consists of discrete steps. Information about each simulation step is stored into the MySQL database. This information is read by the EDU Sandbox utility. This utility is able to learn patterns beforehand. EDU Sandbox reads in real-time data from the database and reports back to NetLogo if any known patterns have been recognized. Based on the recognised pattern, the adaptation logic in NetLogo adapts the chosen action parameters to mediate simulation’s behaviour.

As we can see (Figure 3) Gershenson developed different strategies of self-organising traffic lights. Each strategy has its advantage and disadvantage. The most significant difference is with about 1000 vehicles. The sotl-platoon strategy begins to be ineffective, while the sotl-phase strategy tends to be better for more than about 1000 vehicles. Gershenson’s work does not implement any adaptation strategy where it would be possible to switch between strategies based on the current traffic status. Our experiment tries to do that.

We use one output indicator called Average Speed of Cars which tells us what the average speed of all vehicles in the simulation is. The goal will be to allow the vehicles to pass through this grid system as fast as possible regardless the traffic density.

Desired value of output indicator

For our purpose, there is one key indicator which is Average Speed of Cars. We require that its value would be as high as possible.

Action parameters and their range

For the choice of action parameters (parameters which can effectively change the system’s behaviour; traffic lights strategy parameters, etc.) it is crucial to define their range where behaviour of the output indicator is desirable (Table 1). It includes the minimum and maximum (or enumeration) values of the output indicator given its stability and possible oscillations as well.

Figure 3. Traffic lights SOTL strategies comparison

Download as

Veiw figure View Figure

Simulation

In this phase we have to conduct a necessary amount of experiments with action parameters’ candidates and their range to observe characteristic system’s behaviour observable on the output indicator.

Pattern and expectation determination

In this step, we should have characteristic behaviour patterns split into several categories (Figure 4). Each category represents one type of behaviour and an anticipation what would happen after that behaviour emerges. In (Table 1) we have two categories with corresponding anticipations.

Figure 4. Patterns determination

Download as

Veiw figure View Figure

Table 1. Font Sizes for Papers

Download as

View current table in a new window

Tables index

Veiw figure View Table

Neural network adaptation

In this step, all patterns are presented to the neural network to learn them. Here are some examples of training sets (Figure 5).

Figure 5. Category 1 patterns example

Download as

Veiw figure View Figure

Behaviour Pattern Recognition

When the neural network has been adapted to recognize all training patterns, we can launch the simulation again and see how patterns are recognised. In our case, the recognized patterns appear in our EDU Sandbox utility and are transferred via MySQL to NetLogo according to adaptation logic.

Adaptation of Parameters

In case we had chosen correct patterns and their adaptation had been held successfully, we can adapt the chosen input action parameters of the simulation using application logic implemented in NetLogo. The system should behave according to our expectation. In our case, it should switch between two strategies in order to allow the vehicles to pas as fast as possible (Table 1).

Figure 6 shows the experiment’s result. First, original experiments for solt-phase and sotl-platoon control mechanism were repeated. Two other lines, pattern3vzory4_02 and grid_2def_3pattern represent our adaptation strategies, each with slightly different patterns. These two strategies adaptively combine both original strategies sotl-phase and sotl-platoon based on the traffic density.

Figure 6. Results of adapting system’s behaviour

Download as

Veiw figure View Figure