Swarm-Based Multi-agent Simulation: A Case Study of Urban Traffic - - PowerPoint PPT Presentation

Swarm-Based Multi-agent Simulation: A Case Study of Urban Traffic Flow in the City of Wroclaw

Kr´

• l, D. & Mro˙

zek, M.

ICCCI 2011

2011

Problem

Big Cities, Big Problems

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 2 / 20

Problem

Big Cities, Big Problems : Traffic Congestion

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 2 / 20

Overview

◮ The Kr´

• l & Mro˙

zek Model (2011).

◮ Ant Algorithm; why and how. ◮ Experiment 1: Simulating traffic flow. ◮ Experiment 2: Using the Ant Algorithm.

◮ Future Research. ◮ Alternative Model. ◮ Discussion.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 3 / 20

The Goals

Goals:

◮ Developing a model of a road traffic environment. ◮ Use this model to optimize traffic flow.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 4 / 20

The Goals

Goals:

◮ Developing a model of a road traffic environment. ◮ Use this model to optimize traffic flow.

Focussing on:

◮ National and regional highway networks passing through a city. ◮ Movements of individual vehicles. ◮ Timing of traffic lights.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 4 / 20

The Model

Traffic flow model, consisting of

◮ Roads and traffic lanes. ◮ Intersections. ◮ Traffic lights. ◮ Vehicle density.

City of Wroclaw

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 5 / 20

Limitations

In order to simplify the mode, and to make it more general, the following limitations were set:

◮ Only the main roads and intersections are included. ◮ Roads allow traveling into both directions. ◮ Only traffic that is passing though the city is considered.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 6 / 20

Agent Restrictions

Agents are:

◮ Vehicles. ◮ Intersections.

Several restrictions in vehicle agent behaviour ensure a realistic and dynamic model:

◮ Uphold the local speed limit. ◮ Speed is dependent on the

distance to the nearest

• bstruction.

◮ Acceleration and deceleration is

distributed non-uniformly. distsafe[m] = v[kmh−1]

10

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 7 / 20

Parameter Tuning

Prior to experimenting, to ensure ecological validity: Calculate and compare the traffic density of the model with data from the real traffic survey. if ρsurvey > ρprediction, then generate more vehicles. if ρsurvey < ρprediction, then generate less vehicles.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 8 / 20

Experiments

Two experiments:

◮ Determine the influence of adding a new section of road on the traffic

density.

◮ Determine the influence of the ant algorithm system.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 9 / 20

New Transit Route

City of Wroclaw Measure the difference in traffic density per hour for the model with, and the model without the additional section or road.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 10 / 20

New Transit Route Cont’d

Results On average a 34% reduction in the traffic density. However, this reduction is unevenly distributed with respect to the time of day.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 11 / 20

The Ant Algorithm

Each vehicle drops pheromones while waiting and while crossing an

• intersection. Therefore, the amount of pheromones on an intersection is

related to the congestion around it and the pheromone coefficient; ρph. The amount dropped per one vehicle is : ∆τ =

1 tw ρph

These pheromones also evaporate, which is related to the time difference and to the evaporation rate; ρev : τt = τt−1 − (tt − tt−1)ρev

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 12 / 20

The Ant Algorithm Cont’d

The mean pheromone level is taken as a baseline. The higher the positive deviation from this mean, the sooner a vehicle is inclined to change route, and the longer a corresponding traffic light stays green. Similarly, a higher negative deviation will result in not changing the route, and in a traffic light staying red much longer.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 13 / 20

The Ant Algorithm Cont’d

To optimize the traffic flow, tune:

◮ Pheromone coefficient, ρph ◮ Pheromone evaporation rate, ρev

If either the coefficient is set too high, or the evaporation rate is set too low, then the pheromones will pile up and the system will work chaotic. In the opposite case, no pheromones will be left for agents to observe and thus the system will have no effect.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 14 / 20

The Ant Algorithm - Results

Results On average a 12% reduction in the traffic density. However, here too this reduction is unevenly distributed with respect to the time of day.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 15 / 20

The Ant Algorithm Model - Results Cont’d

Some inferences that can be made from these graphs:

◮ The traffic density on the road is related to the time of day. ◮ The algorithm can effectively handle variations in traffic density. ◮ Static (or sub-optimal tuned) parameters for the pheromone

coefficient and evaporation rate only result in a sub-optimal performance. Pheromones build up and stayed too long, thus guiding traffic away from quiet intersections. In addition, traffic lights gave quiet roads too long a slot.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 16 / 20

Conclusion

The experiments demonstrated that

◮ the model proved very effective in simulating road traffic movements,

and

◮ it proved to be a suitable model for traffic flow optimisation, such as

analysing changes to the road network. Additional benefits of the implementation are:

◮ It can be easily customised to fit most road networks. ◮ It is easily scalable to run on multiple servers.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 17 / 20

Future Research

Several drawbacks were found that might be addressed in future research. These are :

◮ Difficulties with specific road configurations.

All local roads and less important highways were omitted, as well as small intersections and those intersections without traffic lights. This design choice might have had a profound impact on the final dynamics of the model.

◮ Allow for more complex(dynamic?) road events, such as sudden

traffic jams or road-works.

◮ Update the data more often as to allow for a more realistic setting.

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 18 / 20

Alternative Model

Distributed route planning with delegate multi-agent systems over an ad-hoc network.

Tycho Bismeijer, Xander Wilcke and Martin Stolk, Emergence research group, Behavior Dynamics 2011, VU

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 19 / 20

Discussion Discussion

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 20 / 20

Discussion Discussion

Points of discussion :

◮ In a real-life situation, how can pheromones be implemented? What

collects them? How do they return to the vehicle?

◮ How to avoid a sinus-wave type of traffic density? ◮ Would drivers listen to the system, e.g. when the detour is notacibly

longer?

Xander Wilcke (wex.wilcke@few.vu.nl) Swarm-Based Multi-agent Simulation October 2, 2012 20 / 20