You are on page 1of 2

Traffic Flow Prediction with GNP

Paper:

Traffic Flow Prediction with Genetic Network


Programming (GNP)
Huiyu Zhou, Shingo Mabu, Wei Wei, Kaoru Shimada, and Kotaro Hirasawa
Graduate School of Information, Production and Systems, Waseda University
Hibikino 2-7, Wakamatsu-ku, Kitakyushu, Fukuoka 808-0135, Japan,
E-mail: {zhy836@toki., mabu@aoni., weiwei@fuji., k.shimada@aoni., and hirasawa@}waseda.jp
[Received December 8, 2008; accepted April 3, 2009]

In this paper, a method for traffic flow prediction


has been proposed to obtain prediction rules from
the past traffic data using Genetic Network Programming (GNP). GNP is an evolutionary approach which
can evolve itself and find the optimal solutions. It has
been clarified that GNP works well especially in dynamic environments since GNP is consisted of directed
graph structures, creates quite compact programs and
has an implicit memory function. In this paper, GNP
is applied to create a traffic flow prediction model.
And we proposed the spatial adjacency model for the
prediction and two kinds of models for N-step prediction. Additionally, the adaptive penalty functions are
adopted for the fitness function in order to alleviate
the infeasible solutions containing loops in the training process. Furthermore, the sharing function is also
used to avoid the premature convergence.
Keywords: traffic, prediction, GNP

1. Introduction
Nowadays, traffic navigation systems are becoming important parts of the car society. Almost every car is
equipped with an automotive navigation system in many
developed countries. As GPS navigation systems become
more and more popular, many people are discovering fun
and they are in desperate need of using real-time traffic
information services. However, the traditional navigation
systems using static map and geo-information have difficulty in realizing the global optimum, which requires both
of spacial and temporal information.
Due to the prevalence of broadband wireless equipments, each vehicle can send the real-time traffic information to a centralized computing center, where the optimized navigation guidance is made. The challenge is how
to build a real-time system that can handle such massive
amount of dynamic and complicated traffic data and how
to make reasonable car guidance systems that can achieve
a global optimal system, meaning the least traffic congestion. The proper prediction of the traffic flow is an essential component of any advanced traffic control system and
one of the pillars of advanced management of dynamic
traffic networks.
Vol.13 No.6, 2009

1.1. Background
Intelligent Transportation Systems (ITS) has been studied and developed widely. Especially Advanced Traveler
Information System (ATIS), i.e., the main part of ITS,
provides the real-time traffic information to travelers, thus
help to find alternative routes, which would reduce the delays caused by both incidents and congestions. A wide
variety of techniques for the prediction of traffic volumes
based on the framework of ATIS [1] have been proposed
aiming at solving congestion problems.
Many methods are based on the traffic congestion
prediction like regression analysis based methods, however, Lingrasa and Osborne [2] found that auto-regression
method does not show the stable robustness as the amount
of noise increases in the training dataset.
According to Wardrop [3], every traveler tries to minimize their own traveling time. This selfish routing mechanism leads to the user equilibrium (UE). Further more, in
real cases, travelers may perceive a different travel time
over the same road. Then, the selfish routing mechanism will results in the stochastic user equilibrium (SUE).
However, in order to achieve the System Optimal (SO),
total system travel time have to be minimized [4]. The
UE and SUE are not so efficient in SO situation due to its
user-based selfish aspect.
In particular, Dynamic Traffic Assignment (DTA) [5]
has attracted recent attention, because it is capable of processing time-varying properties of traffic flow. However,
these complex formulations generally lead to extremely
complicated solutions and another issue is that the requirement of time-dependent OD (original and destination) data in prediction process.
Neural Networks (NN) [6] have also been used to traffic congestion prediction. Unlike other methods, NN depends on the self-evolved parameters and recurrent calculation under the teacher signal more than the complicated mathematical analysis, thus it is easier to process
and also shows better robustness under the noisy environment [2]. Whats more, NNs have a good adaptive ability
since the parameters can be adjusted automatically as the
OD changes in the environment.
A new evolutionary method is proposed in this paper.
Genetic Network Programming (GNP) is used to construct the basic model for traffic volume prediction and

Journal of Advanced Computational Intelligence


and Intelligent Informatics

713

Zhou, H., Mabu, S., Wei, W., Shimada, K.,and Hirasawa, K.

the structure of it is self-evolved until we get the optimal global structure. Thus, the proposed method also has
the robustness and adaptive ability shown in NN. Furthermore, the features of GNP make it possible to better handle the problem of local minima NN is facing.
Another feature of the proposed method is that it does
not need to use the time-dependent OD data like some
other ATIS based methods. After the traffic volume is
observed, the evolutionary method GNP is applied to selfevolve and the prediction model is built automatically, and
after the training phase finishes, the generated best individual in GNP would become the prediction model.

1.2. Feature of Proposed Method


One of the evolutionary approach-Genetic Network
Programming (GNP) [79] has been adopted to build the
prediction model, to train it and to obtain the prediction
rules, because evolutionary approaches have the following advantages when working on the complicated optimization problems:

Firstly, GA-like evolutionary approaches are intrinsically parallel. Most other algorithms are serial and
can only explore the solution space to a problem in
one direction at a time.

Due to the parallelism that allows them to implicitly


evaluate many schema at once, they are particularly
well-suited to solving problems, where the space of
all potential solutions is truly huge, that is, too vast
to search exhaustively in any reasonable amount of
time.

Another strength is that they perform well in problems for which the fitness landscape is complex, discontinuous, noisy, changes over time, or has many
local optima.

Another one is their ability to manipulate many objectives simultaneously. Many real-world problems
cannot be stated in terms of a single objective to be
minimized or maximized, but must be expressed in
terms of multiple objectives, usually with tradeoffs
involved: one can only be improved at the expense
of others.

In addition, some mechanisms have been used to improve the performance, i.e., the penalty function and sharing function, when applying GNP to the prediction of the
traffic. The features of the traffic flow prediction with Genetic Network Programming (GNP) are:

Spatial adjacency model is introduced, which uses


neighboring sections related to the objective section
firstly and secondly for the traffic flow prediction
with GNP is built. And two kinds of N-step traffic
flow prediction methods are used in this paper.

Adaptive penalty functions [10, 11] are added to the


fitness function to inhibit the infeasible solutions

714

containing loops in the training, as a result, the solutions become self-adapted towards the direction of
more legal individuals;

The third one is to use the sharing function [12] to


avoid the premature convergence.

2. Genetic Network Programming (GNP)


GNP is an graph-based evolutionary approach extended
from the Genetic Algorithm (GA) [13, 14] and Genetic
Programming (GP) [15, 16].
GA proposed by J.H. Holland, has also been applied
to many areas. However, because of string representation
of GA, it is difficult to obtain a complete picture of the
underlying relationships in problem domains.
GP improved the expression ability of GA by evolving individuals as tree structures. Although, GP enables
a more explicit representation of reference rules, it may
suffer from the problem of loose structure and bloating,
especially in dynamic problems. GNP inherited some
good characteristics such as global behavior, easy to define the underlying representations, etc, from GA and GP.
Besides, the distinguishing features of the GNP structure
are the directed graph expression, reusability of nodes,
and implicit memory function. In the concrete, the features of GNP are like: 1) GNP programs are composed of
a number of nodes which execute simple judgment/ processing, and these nodes are connected by directed links
to each other. 2) The graph structure enables GNP to reuse nodes, thus the structure can be very compact, furthermore, the directed graph expression can realize some
repetitive processes. 3) The node transition of GNP is
executed according to its node connections, thus the past
history of the node transition affects the current node to be
used and this characteristic works as an implicit memory
function.
The basic structure (phenotype) of GNP [79] is shown
in Fig. 1. GNP is composed of three kinds of nodes: a
start node, judgment nodes and processing nodes. Judgment nodes correspond nearly to elementary functions of
Genetic Programming (GP). Judgment nodes are the set
of J1 , J2 , . . . , which work as if-then type decision making functions. On the other hand, processing nodes are
the set of P1 , P2 , . . . , which work as some kind of action/processing functions. The practical roles of these
nodes are predefined and stored in the library by supervisors. Once GNP is booted up, firstly the execution starts
from a start node; secondly the next node to be executed
is determined according to the connection from the current activated node or the judgment result in the current
judgment node.
The genotype expression of GNP node is shown in
Fig. 2, which describes the gene of node i. The set of
these genes represents the genotype of GNP individuals.
NTi describes the node type, NTi = 0 when node i is the
start node, NTi = 1 when node i is a judgment node and
NTi = 2 when node i is a processing node. IDi is an iden-

Journal of Advanced Computational Intelligence


and Intelligent Informatics

Vol.13 No.6, 2009

You might also like