Flexible Multicriterial Agenda Planning in Public Transit Systems

An Intelligent Agent for Mobility-oriented Agenda Planning

Felix Schwinger

, Fabian Ohler

and Karl-Heinz Krempels

1,2

Fraunhofer Institute for Applied Information Technology FIT, Aachen, Germany

Information Systems, RWTH Aachen University, Aachen, Germany

Keywords:

Agenda Planning, Information System, Intelligent Agent, Mobility Planning, Operations Research, Optimiza-

tion Problem, Orienteering Problem, Public Transportation, Travel Planning.

Abstract:

Planning a mobility-oriented agenda is a time-consuming and tedious task for many travelers. A person is

required to collect information from different sources such as a map service, a business register, a calendar and

a journey planner. However, she is mostly not interested in either planning the agenda or the journeys between

different locations of the agenda; but is more interested in completing the tasks of the agenda. Therefore, we

propose an intelligent agenda planning agent that aims to support people with this task. We integrate public

transit schedules with additional spatial information from OpenStreetMap to create an information database

for the agent. The agent can then plan tasks and appointments and the mobility between those items. First

brief evaluations with a survey have shown, that the algorithm ﬁnds shorter agendas than most manually found

agendas. However, participants of the survey criticized the temporal placement of tasks in the agenda.

1 INTRODUCTION

Urban transportation systems are often large, com-

plicated to use for the ﬁrst time and when regarding

multimodal transportation systems, also difﬁcult to uti-

lize effectively (Zografos and Androutsopoulos, 2008).

In recent years, there has been a surge of new infor-

mation systems that provide useful information for

passengers of public transit systems. The developed

information resources include map services, business

directories, and online route planners which try to de-

crease the usage complexity of urban transportation

systems. A user can resort to these services when

planning her agenda; however, the main problem with

these services is that they do not provide an easy to

understand, integrated solution to the whole problem

of mobility agenda planning, and only offer segmented

information (Baena-Toquero et al., 2014). A mobility

agenda does not only include the agenda items, i. e.,

appointments and tasks, but also the mobility between

the different locations of the agenda. The user, there-

fore, needs access to multiple services when planning

her agenda: She could check nearby locations on a

map service, then inquire their opening times in a busi-

ness directory and ﬁnally plan the journeys between

the locations with a route planner. Additionally, she

might check different journey planners for different

types of mobility, i. e., car-sharing, walking or pub-

lic transportation. A user might, however, not even

be interested in manually planning her agenda, but is

more interested in an agenda that allows her to per-

form all tasks she wants to complete on a day. An inte-

grated intelligent system that allows the user to specify

her needs, which then calculates an optimal mobility

agenda that ﬁts the passenger’s schedule would help

to drastically reduce the complexity. This paper intro-

duces an algorithm for computing intra-city itineraries

that focuses on public transportation systems.

The remainder of this paper is structured as fol-

lows: In Section 2, we propose the Mobility Agenda

Planning Problem and in Section 3, the related work

is introduced and discussed. Section 4 introduces a

model for mobility-oriented agenda planning and Sec-

tion 5 describes an algorithm for it. We show ﬁrst brief

results of a performed survey in Section 6 and ﬁnally

conclude the paper with an outlook and a summary in

Section 7.

2 PROBLEM DESCRIPTION

In this paper, we regard the Mobility Agenda Planning

Problem (MAPP) as follows: The input for the prob-

558

Schwinger, F., Ohler, F. and Krempels, K.

Flexible Multicr iterial Agenda Planning in Public Transit Systems.

DOI: 10.5220/0006795305580567

In Proceedings of the 4th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2018), pages 558-567

ISBN: 978-989-758-293-6

lem is a number of appointments and tasks. Appoint-

ments are entirely ﬁxed, whereas tasks are movable

in at least one dimension, i. e., temporally or spatially.

Each insertable agenda item has an associated score

and a completion time. The score of the agenda item

may represent the priority of the task to the person,

or a combined score of various factors. The goal of

a MAPP algorithm is to ﬁnd an optimized mobility

agenda, containing all appointments and a number of

tasks. Additionally, the algorithm computes the mobil-

ity between all locations in the agenda. In this paper,

we focus on walking and public transportation jour-

neys between agenda items. The task for the problem

is then to optimize the mobility agenda by maximiz-

ing the score of the agenda while minimizing the total

time it takes to complete the agenda. The total time

includes possible wait times at public transit stations,

wait times at Points of Interest (POIs), travel times be-

tween different locations, and the completion time for

agenda items. All appointments have to be reached on

time, and all public transit transfers must be feasible.

Furthermore, the algorithm has to consider category

or location constraints for agenda items and potential

dependencies between agenda items. Each location,

therefore, also has an associated category, so that the

agent can determine suitable locations for tasks. It

should also be possible to deﬁne dependencies be-

tween agenda items so that one agenda item can only

be inserted into the agenda, if another agenda item

has been completed before it. To compute mobility-

oriented agendas, the algorithm needs to optimize the

spatial placement of tasks and their sequence. To this

end, the algorithm needs to know the location of POIs

including their category and be able to generate routes

in a road network, as well as in a public transit net-

work. The agent can then compute a mobility itinerary,

containing a list of agenda items and a travel plan on

how to reach the different location either by walking,

by public transit, or by a combination of both travel

modes.

This work aims to investigate whether it is possible

to create an intelligent agent to support people in the

agenda planning process and combine the process with

mobility planning in a real-life setting.

3 RELATED WORK

Not much research has been conducted on general

mobility-oriented agenda planning. The research on

combining mobility and agenda planning also incor-

porating public transportation has, to the best of our

knowledge, only recently started. Other optimization

problems such as the Tourist Trip Design Problem

(Vansteenwegen and Van Oudheusden, 2007) are, how-

ever, a specialization of the Mobility-Agenda Planning

Problem, which focuses on tourist agendas.

3.1 Model of Mobility Agenda Planning

A model for mobility agenda planning has been pro-

posed by (Wienken et al., 2014; Wienken et al., 2017).

The authors discuss the possibility of combining mobil-

ity planning and agenda planning. This integration is

possible due to a large similarity between both aspects.

Table 1: Similarity between information needs for agenda

planning and mobility planning. (Wienken et al., 2014)

Information need of

Agenda planning Mobility planning

Purpose Location

Date/Time Date/Time

Location Sequence of mobility

Sequence Means of transportation

Social Mobility impairments

Table 1 shows the identiﬁed information needs.

The table indicates that both agenda items and mo-

bility elements share some information needs such as

the location and the time. A mobility agenda con-

sists of agenda items and the mobility between the

different agenda items. An agenda item has a purpose

which determines whether it is a task or an appoint-

ment. The problem for the mobility agenda planning

agent is to compute a mobility-oriented agenda that

allows the user to be on time to all her appointments

while performing an optimal number of tasks at their

respective locations. Between the agenda items, the

mobility agenda planner should also plan the mobility

with different choices of transportation. The algorithm

may tailor the selection of transportation and agenda

items to a speciﬁc user proﬁle, respecting the user’s

preferences.

As an example, a user gives the agenda planner

a list of ﬁxed appointments on that day and a list of

tasks she wants to perform that day. The function of

the mobility agenda planner is then to ﬁnd a suitable

location for the tasks if a location for the task is not

speciﬁed. The mobility agenda planner should choose

places in such a way that the resulting mobility agenda

is optimal to the user, for example, as short as possible,

while still allowing for the completion of all tasks.

3.2 Relevant Optimization Problems

In the literature, we have not found optimization prob-

lems that deal with computing mobility-oriented agen-

das in a general way, but we have found several more

Flexible Multicriterial Agenda Planning in Public Transit Systems

559

specialized problems that can be generalized and be

used as a basis for a mobility agenda planning algo-

rithm.

Tourist Trip Design Problem.

The MAPP is a gen-

eralization of the Tourist Trip Design Problem (TTDP)

(Vansteenwegen and Van Oudheusden, 2007). The

TTDP is a route planning problem that focuses on

tourists, which want to visit multiple POIs in a city.

The objective of the TTDP is to select a subset of POIs

that match the user’s proﬁle and to maximize the user’s

satisfaction with the route. The authors suggested to

calculate the satisfaction of a user with various param-

eters and constraints such as the choice of POIs, the or-

der of the visits, the distance between POIs, the visiting

time for POIs, the opening hours of POIs and further

information such as entrance fees or weather condi-

tions. The main difference between the TTDP and the

MAPP is that the TTDP does not differentiate between

appointments and tasks. Furthermore, the TTDP does

not necessarily model dependencies between tasks and

mostly does not distinguish between various categories

of POIs. The mobility-oriented agenda planning agent

must place certain tasks of the user, i. e. buying bread,

at certain locations, such as a supermarket or a bakery,

which is why the algorithm needs to distinguish be-

tween different categories and must map the intent of

the user to certain POI categories. For a MAPP query

it should, furthermore, be possible to plan certain tasks

not before other tasks have been completed, e. g. visit

the pharmacy after visiting the doctor.

Orienteering Problem.

The TTDP has often been

modeled as an Orienteering Problem (OP) in the lit-

erature (Vansteenwegen and Van Oudheusden, 2007).

The OP, also known as the Selective Traveling Sales-

man Problem, can be seen as a combination of the

Traveling Salesman Problem (TSP) and the Knapsack

Problem (KP). In the OP we have a graph with vertices

and arcs. The vertices represent POIs, while the arcs

represent the travel time between POIs. Each vertex

is annotated with a score and a visiting time, and each

arc is annotated with a travel time. The task of the

OP is to compute a path visiting a certain subset of

vertices in such a way, that the aggregated score of the

vertices is maximized. The path is constrained by the

time budget and the visiting times of the vertices added

to the travel times on the arcs is not allowed to exceed

the time budget. Furthermore, the score of each vertex

is only allowed to be collected once. (Golden et al.,

1987) have proven that the OP is NP-hard. Hence, it is

infeasible to calculate exact results for larger problem

instances and heuristics need to be found to determine

quality solutions to the OP. Nearly all recent papers

published on the OP and its variations describe an

approximation algorithm for the problem.

More advanced approaches model the TTDP as

a Team Orienteering Problem with Time Windows

(TOPTW) (Boussier et al., 2007). The TOPTW is

an extension to the OP, where the task is not to ﬁnd a

single route in the graph that maximizes the total score,

but

routes, which maximize the aggregated score of

all routes. Each route has a time budget and the score

of a route is the score of the route’s POIs. The score

of each vertex is only allowed to be collected once.

This extension of the OP was used to allow comput-

ing tourist trips spanning multiple days. Additionally,

the OP has been modiﬁed to support time windows

of vertices, so collecting the score of a vertex is only

possible, if it is visited during its time window. With

this extension, we can model the opening and closing

times of POIs in the TTDP.

Furthermore, our approach should integrate public

transportation as a means to travel. Later, we want to

extent this support to new mobility offers such as car-

or bike-sharing. Public transport, however, is schedule-

based and therefore the travel times from one location

to another are time-dependent. Depending on when a

person arrives at the bus station, the travel time may

differ. For dealing with time-dependency, the Time-

Dependent Team Orienteering Problem with Time Win-

dows (TDTOPTW) has been proposed. The problem

description itself is the same as for the TOPTW, but

the travel times on the arcs of the graph are now a

function, whose value depends on the time for which

it is traversed. Computing exact solutions, even for

small problem instances of the Time-Dependent Ori-

enteering Problem, is not feasible (Gunawan et al.,

2014). The problem was attempted to be solved with

the commercial Integer Problem (IP) solver CPLEX,

which was not able to compute exact solutions to small

problem instances in less than 24 hours. As interac-

tions with an intelligent agent should be possible in

nearly real-time, the runtime of the algorithm should

be limited to less than a few seconds, making exact

solutions not feasible for our use case.

For solving a TDTOPTW there are two main ap-

proaches discussed in the literature: The key idea

of the ﬁrst approach is to reduce a TDTOPTW to a

TOPTW by assigning each arc in the graph the aver-

age travel time from one location to another (Garcia

et al., 2013). Removing the time-dependency from the

problem reduces its computational complexity. The

TOPTW is then solved with an Iterated Local Search

(ILS) algorithm. An ILS algorithm is a meta-heuristic

algorithm applicable to a wide range of optimization

problems (Louren

c¸

o et al., 2003). It operates in several

iterations and performs local search steps in each iter-

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

560

ation in an attempt to improve the best-found solution

so far. While this approach reduces the runtime of

the algorithm, the downside is that it may return infea-

sible mobility agendas, because it does not consider

the actual schedule of the public transit vehicles. To

handle these cases the authors in (Garcia et al., 2013)

implemented a repair procedure, which uses a time-

dependent route planner, to validate the solution. If an

error is found, it attempts to repair the created route

by removing visits from it.

The second approach also solves the TDTOPTW

with an ILS algorithm (Gavalas et al., 2015). The

key difference is, that this solution directly integrates

the time-dependency and the problem is not reduced

to a TOPTW, but is rather directly solved. The au-

thors improved the runtime of the algorithm by using

a clustering approach, which limits the search space of

the algorithm and therefore shortens its runtime. The

evaluation showed that this clustering heuristic only

minimally affects the solution’s quality, but greatly

improves the runtime, when compared to the default

ILS heuristic.

Finding Optimal Sequences.

Another aspect of

mobility agenda planning is to ﬁnd optimal sequences

of POIs. For this, we have researched optimization

algorithms. The Trip Planning Query (Li et al., 2005),

Optimal Sequence Route (Sharifzadeh et al., 2008),

and Multi-Request Route Planning (Lu et al., 2017)

problems all deal with ﬁnding a sensible sequence for

visiting a number of POIs. They do, however, not in-

clude the functionality to reduce the requested POIs to

a subset, but rather plan an optimal route containing

all requested task, similar to the Traveling Salesman

Problem. This means that there is no notion of a time

budget, which the algorithm could consider. Addition-

ally, as there is no time budget, there is also no score

of the POIs, the algorithm rather minimizes the travel

time required for the route, but does not consider a kind

of score for the various POIs. Furthermore, these ap-

proaches are designed for road networks, meaning that

the travel times are assumed to be time-independent

and symmetric, which is both not the case for a public

transit network.

4 MODELING OF THE PROBLEM

We propose to model the MAPP as a TDTOPTW and

transform it the following way: As appointments are

entirely ﬁxed an inﬂexible, we can construct a corre-

sponding TDTOPTW problem, where appointments

are already handled. A graphical illustration is shown

in Figure 1. Step

creates an empty solution with the

starting and ending location of the agenda with the re-

spective starting and ending times. In Step

we create

a new route for each appointment in the agenda. This

means, that we model the time of an appointment as

the time difference between two routes. This modeling

ensures, that the TDTOPTW algorithm respects the

time windows and locations of the appointments. For

example, the complete mobility agenda of a user starts

at 8:00, ends at 18:00, and there is an appointment

from 14:00 to 16:00. We model this problem as a TD-

TOPTW with two distinct routes: The ﬁrst route

starts at 8:00 and ends at 14:00, whereas the second

routes r

starts at 16.00 and ends at 18:00.

Mobility Agenda

Start of the Agenda

End of the Agenda

Empty

(a) Step 1

Mobility Agenda

Start of the Agenda

End of the Agenda

Appointment

Route 1

Route 2

(b) Step 2

Mobility Agenda

Start of the Agenda

End of the Agenda

Appointment

Transit Time

Task 1

Transit Time

Task 2

Transit Time

Figure 1: Creation of a TDTOPTW instance from a MAPP

instance, consisting of two tasks and one appointment. The

corresponding TDTOPTW therefore consists of two routes.

The constructed problem ensures that the user is on

time to her appointment because route

ends at the

location of the appointment at the time of the appoint-

ment. For each appointment of the user, we therefore,

end a route, at the time when the appointment begins

and start a new route, for the time when the appoint-

ment ends. The respective ending and starting loca-

tions of the routes are the locations of the appointments.

As the CSCRoutes algorithm only inserts a task, if no

route time window is violated, the timely arrival at the

appointment location is guaranteed. After inserting the

appointments into the solution, the algorithm can start

to add tasks into the routes. The algorithm then has

two routes available to insert tasks into. The insertion

of tasks is performed in Step

. Step

is done with the

modiﬁed CSCRoutes algorithm, while Step

and

performed by a pre-processing algorithm constructing

the associated TDTOPTW problem. The CSCRoutes

algorithm is introduced in more detail in Section 5.2.

The modiﬁcations include the support of an order of

agenda items and category constraints on the POIs.

Flexible Multicriterial Agenda Planning in Public Transit Systems

561

We identiﬁed various other features that may be

important for an agenda planner to support such as:

Mutual exclusion of agenda items, deadlines, preferred

order of agenda items, mandatory POIs (not an appoint-

ment), time-dependent scores for POIs, multimodality,

and connection between agenda planners (i. e. plan a

task together with other people). Some of them can eas-

ily be integrated into our approach such as deadlines,

for example; while the integration of multimodality is

more complex.

Scoring of Agendas.

The main task of the agenda

planner is to help the user plan tasks on a speciﬁc day.

We have two ratings that are important for an agenda

plan, the score of the agenda, denoting the preference

to the user and the completion time of the agenda, i. e.,

the time it takes for the user to complete the agenda.

Both of these ratings have to be correlated and be used

for rating the agenda, otherwise unreasonable agendas

may be returned. We have decided to use a weighted

linear combination of the score of the agenda and the

completion time. The score of the individual agenda

tasks, however, must also be correlated to the time

taken to complete them, otherwise the agenda planner

may prefer low-scored short tasks over high-scored

long tasks. Furthermore, the score of the individual

agenda tasks is also an combined score consisting of

several factors such as the priority of the task, the

preference of the user and maybe further external in-

formation such as the weather forecast, for example.

A user may be more inclined to perform tasks outside,

for a nice day and prefer doing indoor activities on a

rainy day. When using the weather forecast as input

for the score of an agenda item, the score must also be-

come time-dependent, as it may change during the day.

The weather forecast may also inﬂuence the mobility

choice of the user i. e. prioritize the bus on rainy days.

5 AGENDA PLANNING

ALGORITHM

We compute mobility-oriented agendas with the help

of a modiﬁed version of the Cluster Search Cluster

Routes (CSCRoutes) algorithm proposed in (Gavalas

et al., 2015). These modiﬁcations include the addi-

tion of various categories based on the OpenStreetMap

category model, handling of dependencies between

various insertable locations, computing the journey

times between various locations on-the-ﬂy during the

execution of the algorithm and improving the algo-

rithm’s runtime by designing further insertion heuris-

tics. CSCRoutes is a heuristic algorithm for the TD-

TOPTW based on the ILS meta-heuristic.

5.1 Workﬂow

The mobility-oriented agenda planning agent requires

data from multiple information sources as it integrates

agenda planning with mobility planning: A mobility-

oriented agenda planner needs to have access to POI

information because we require the POI’s name, its

category, its location, its opening hours, and a unique

identiﬁer. The POI information is used for the plan-

ning of tasks, as they may need to be performed at

special locations, i. e. buying bread at a bakery. Fur-

thermore, the algorithm must be able to compute pub-

lic transit and walking journeys between the locations

of different agenda items. To this end, we also need

the schedule of the public transportation service. The

workﬂow of the algorithm, which is illustrated in Fig-

ure 2, is then as follows: We ﬁrst obtain the POI and

road information from OpenStreetMap and extract

the relevant information from it. Then, we also ob-

tain the relevant public transportation information as a

GTFS ﬁle

and combine it with the OpenStreetMap ex-

tract using Graphhopper to obtain routing information.

Graphhopper is an open source route planning library,

which recently added support for public transit plan-

ning. The CSCRoutes algorithm queries Graphhopper

during the execution and no routing information be-

tween POIs is pre-computed. In contrast to that, the

POI database is pre-computed by the algorithm from

the OpenStreetMap extract. Afterward, we use the POI

information and the route planner to run the adapted

version of the CSCRoutes algorithm (Gavalas et al.,

2015).

5.2 Optimization Algorithm

In the ﬁrst step of the MAPP algorithm, before execut-

ing the optimization algorithm, we transform it into a

TDTOPTW as explained in Section 4. The transforma-

tion already ensures, that all appointments are correctly

planned. An optimization algorithm is then executed

on the remaining spare time in order to ﬁnd suitable

tasks for this spare time. The optimization algorithm

itself is mostly adopted from the CSCRoutes algorithm

(Gavalas et al., 2015), with various changes to adapt

it to the domain of agenda planning. The algorithm is

changed to allow an enforced order on agenda items

and to be able to handle POIs of different categories.

CSCRoutes is based on the Iterated Local Search (ILS)

meta-heuristic from (Louren

c¸

o et al., 2003) and em-

ploys several local search steps. The pseudocode of

the ILS algorithm is shown in Algorithm 1. The ILS

meta-heuristic consists of four methods that are re-

quired to be implemented: GenerateInitialSolution,

https://developers.google.com/transit/gtfs/

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

562

OSM File

GTFS File

Osmosis OSM Extract OsmPoisPbf POI Information

CSCRoutes

Algorithms

Routing InformationGraphhopper

Computed

Mobility Agenda

Figure 2: Workﬂow of the MAPP Algorithm. The rectangles show information, and the ellipses depict a modiﬁcation of

information.

LocalSearch, Perturbation and AcceptanceCriterion.

At ﬁrst, we generate an initial solution with the Gen-

erateInitialSolution method and search for an local

optimum using the LocalSearch method. After we

initialized the solution the algorithm is repeated until

the termination criterion is met. In this loop the algo-

rithm attempts to escape from the local optimum with

the Perturbation method from the previous iteration,

reaches a new local optimum with the LocalSearch

method and selects the starting solution for the next

iteration with the AcceptanceCriterion method. When

the termination criterion is met, the algorithm returns

the best found solution.

Algorithm 1

: Iterated Local Search Algorithm. Adapted

from (Lourenc¸o et al., 2003).

1: s

← GenerateInitialSolution

2: s ← LocalSearch(s

)

3: repeat

4: s

← Perturbation(s, history)

5: s

← LocalSearch(s

)

6: s ← AcceptanceCriterion(s, s

, history)

7: until termination criterion met

The CSCRoutes algorithm implements the differ-

ent methods of the ILS algorithm. CSCRoutes has a

route initialization phase which is run at the beginning

that initializes the different routes with ﬁrst tasks. Af-

terward, local search operations are performed in order

to reach the local optimum. CSCRoutes has two local

search operations implemented, the insertion step and

the replace step. The insertion step inserts new tasks

into the different route, until no insertion is possible

anymore, while the replace step attempts to replace

already inserted tasks with higher scored non-inserted

tasks. Perturbation is implemented in the shake step.

The shake step removes a certain number of tasks from

the routes, so that new tasks can be inserted in the

next iteration. So that the CSCRoutes algorithm does

not get trapped in the same local optimum over and

over again, it always continues its search from its last

found solution, After reaching a pre-deﬁned number

of non-improving iterations, the CSCRoutes algorithm

terminates.

Heuristics.

As the search space for the agenda plan-

ning algorithm is large, especially if it has a nation-

wide database of POIs, heuristics need to be used to de-

crease the algorithm’s runtime. Hence, the CSCRoutes

algorithm only checks certain POIs for insertion in

each iteration. For this, it assigns clusters to all POIs

in the dataset based on their geolocation in a pre-

processing step. The clusters are then used to con-

strain the insertion and replacement step of the algo-

rithm. With the exception of the starting and ending

location, this constraint forbids it to revisit clusters

in a route. This property is called the cluster route

constraint (Gavalas et al., 2015). Once a POI belong-

ing to a cluster

has been inserted into the routes, all

POIs belonging to

can only be inserted into the route,

when it is adjacent to another POI belonging to cluster

c in the route.

5.3 Problems with Existing Approaches

In Section 4, we have seen that the TDTOPTW needs

to be extended to compute useful mobility agendas.

The TDTOPTW lacks the deﬁnition of categories of

POIs and dependencies between agenda items. In

our use case, categories are required to ﬁnd suitable

locations for tasks. Furthermore, CSCRoutes pre-

processes the data in such a way that using a larger

number of POIs is infeasible (Gavalas et al., 2015). As

a agenda planner should not only be usable in a con-

ﬁned area, but preferably work worldwide, the number

of POIs may become large. The CSCRoutes algo-

rithm, however, computes the fastest journey between

all pairs of POIs for every time step and obtains a large

travel-time matrix. (Gavalas et al., 2015) used

1440

time steps per day, one per minute. These time steps

are needed to deal with the time-dependent nature of

public transit journeys. However, the pre-processing

step only needs to store the non-dominated shortest

travel times between each pair of POIs and not a travel

time for each time step. A departure

dep

dominates

another departure

dep

, if it starts later than

dep

, but

arrives at the same time or earlier at the destination

than

dep

. This pre-processing step is very costly, es-

pecially with larger number of POIs. The advantage of

the preprocessing is that it allows to acquire relevant

Flexible Multicriterial Agenda Planning in Public Transit Systems

563

travel times by a simple lookup in the matrix.

However, pre-calculating all travel times is infeasi-

ble, when regarding a larger number of POIs.

Instead of pre-calculating all pair-wise travel times,

we utilize the GraphHopper

routing engine to ﬁnd

journeys from one location to another location. In

theory, the regions to compute mobility agendas in,

can then be signiﬁcantly larger. The major downside

of this approach, however, is that the computation of

point-to-point journeys takes much longer than a ma-

trix lookup. While we only need to read a single value

from the travel-time matrix obtained by preprocess-

ing the data, we need to compute a viable journey

on-the-ﬂy using a multimodal router. The computation

of journeys takes several orders of magnitude longer

compared to looking up the relevant information in the

matrix. Hence, a trade-off between computation time

and storage requirements exists, but the more POIs are

supported, the less attractive the pre-processing step

becomes.

Another signiﬁcant advantage of the GraphHopper

approach is that, we know the locations of transfers

and the waiting time at transfer stations. This infor-

mation is lost, when only storing the travel time and

disregarding the transfers itself. Lastly, determining

the required information for the user when calculating

the route on-the-ﬂy, is signiﬁcantly simpler. A user

expects to have an itinerary which tells her where to

enter and to exit vehicles and which roads she should

take to her location. The travel-time matrix itself only

stores the travel time, therefore, there is a need for a

post-processing step that attempts to recreate the meta-

information of the route. By directly calculating the

journeys with a route planner, we can easily extract

the routing information for the user and exactly guide

her to her locations.

5.4 Additional Heuristics

As the number of POIs that a mobility-oriented agenda

planner needs to consider is much larger than a solver

for the TTDP, we had to implement further heuristics

to the algorithm to reduce the computation time. We

have a trade-off between quality of the solution and the

algorithm’s runtime. The more the heuristic limits the

search space, the worse the computed result will be,

nevertheless we attempted to ﬁnd heuristics that try to

reduce the negative impact on the result’s quality. The

CSCRoutes algorithm spends most of its runtime in

the scoring function of possible insertions, as it needs

to calculate the additional time an insertion would cost.

To score an insertion of a POI

between a POI

and

https://github.com/graphhopper/graphhopper

, two journeys need to be computed: The traveling

time from p

to p

and from p

to p

Most of the time it is not important for the scor-

ing whether the travel time is slightly inaccurate, as

long as the calculated value for the travel time is close

enough to the real travel time. The idea here is to only

approximate the cost of an insertion in an attempt to

minimize the number of journey computations. As we

have already clustered the POIs, we will also use the

clustering step to enable a better use of caching. In the

scoring step, instead of calculating the travel time from

one POI directly to another POI, we will use the travel

time from a POI to the other POI’s cluster centroid.

On the one hand this approach results in more cache-

hits and allows to reduce the number of costly travel

time computations, whereas, on the other hand the

scoring of insertions is less accurate as the travel time

between POIs is only approximated. Nevertheless, the

correctness of the results is ensured, because we will

calculate the actual travel time between the locations,

when inserting the selected POI into the agenda. As

long as we only regard intra-city mobility agendas at

the moment, this heuristic returns reasonable results,

as most POIs are not far away from their respective

centroids. When extending the support to inter-city

mobility agendas, a better heuristic needs to be found.

Another heuristic regards the selection of possible

POI insertions. The solution of (Gavalas et al., 2015)

regarded most POIs for insertion, only limited by the

cluster constraint. Therefore, we propose to limit the

area in which the algorithm is able to select POIs for

insertion from. This is problematic for the general OP,

as the score of a POI and its location are completely in-

dependent. For the MAPP, however, an approximation

like this is slightly less problematic as we assume that

the score of POIs in the same category does not drasti-

cally differ. We restrict the algorithm, when inserting

POI

between POIs

and

to only search for POIs

which are in a certain radius around

, or along

a corridor on the path from

. In this way we

reduce the search space of the algorithm, while still

allowing for many of the most important POIs to be in-

serted. If no suitable POI could be found for insertion,

the search area can be widened. The downside of this

heuristic is, that POIs that are further away, but where

many other highly-scored POIs are located might be

missed, when they are outside of the search area of

the algorithm. This is not a problem for an agenda

with many appointments and only short time windows

for tasks, but becomes a larger problem when a lot of

spare time is available. When more time is available,

it might be more sensible to travel further to reach a

cluster with several highly scored POIs, i. e. a mall,

because the travel time inside the cluster will be small.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

564

If not much time is available in the route, traveling to

a cluster that is far away does not make much sense, as

performing all tasks at that cluster may not be possible

due to the time constraints of the next appointment.

Setting the search distance for POIs, therefore, could

depend on the available spare time in that speciﬁc time

slot in the route.

6 EVALUATION

In the following section, we will evaluate the practical

feasibility of mobility-agenda planning and our agenda

planning agent.

6.1 Methodology

To evaluate the agenda planning algorithm, we have

designed a short survey with which we attempted to

answer the following questions:

•

Which aspects of agenda planning and mobility

planning are important and how do they interact

with each other?

•

Which features of a mobility-oriented agenda are

most important to users?

•

Which features of a mobility-oriented agenda plan-

ning agent are most important to users?

•

How well does the agent planned agenda fare

against manually planned agendas by the partici-

pants?

The survey was structured into four different parts,

for each of the above listed research questions. We

aimed to evaluate mobility agenda planning, both qual-

itatively and quantitatively. Hence, the fourth part of

the survey was interactive in which we have asked

the participants to plan a mobility agenda of a ﬁctive

persona. We have designed a persona and a scenario

as the agenda planner does not have a user interface

as the moment. This allowed us to ask the partici-

pants to plan an agenda for the same query as our

agenda planning agent. We were then able to com-

pare the manually planned solution to the solution

of the agenda planning algorithm. Additionally, we

have asked the participants to evaluate the agent and

manually planned mobility agenda themselves and list

advantages and disadvantages of both agendas.

To ease the effort of planning an agenda for the

respondents, we designed and implemented a journey

planner based on GraphHopper that overlays the rele-

vant POI information on a map, on an interactive web-

site. On the one hand, this simpliﬁed the question for

the participants, as only a single website was required

to be used. While on the other hand, this allowed the

respondents to work on the same data basis as the op-

timization algorithm. The ﬁctive agenda that had to

be planned, consisted of one appointment and seven

tasks. This scenario was chosen, because planning

appointments is not difﬁcult, as they are completely

inﬂexible. The interesting aspect is the temporal and

spatial placement of tasks. Furthermore, two tasks

had to be completed at speciﬁc POIs and in a speciﬁc

time frame, whereas the others could be completed any

time at their respective category. Additionally, we have

added a single dependency between two tasks. The

agenda then consists of two distinct routes. The total

visiting time of all tasks (without the appointment) is

101

minutes, which does not include the journey time

between the various POIs.

6.2 Results

We implemented a prototype agenda planning agent in

Python 3.5.2 with SciPy for the clustering algorithms

and helper functions, joblib for the parallelization of

the algorithm and GraphHopper for computing public

transit journeys. The software was run inside a vir-

tual machine conﬁgured with 8 vCPUs (Intel Xeon

E5-2650 clocked at 2.20GHz) and 8GB of RAM. As

this survey only attempts to show ﬁrst insights into

the topic and is not meant to be representative in any

way, most of the 14 respondents have been students.

For most of the respondents the location of the sce-

nario was not well-known. Unfortunately, the sample

group is too homogeneous for a meaningful qualitative

feedback, while the sample size is too small for an

expressive quantitative feedback.

The short evaluation of mobility agenda planners

showed that most respondents think that an agenda

planner would help them to organize and plan their

day. They expect the agenda planner to be especially

useful in unfamiliar locations or cities, whereas they

only expect limited utility in known places. The mini-

mization of unintended breaks in the agenda and the

completion time was listed as crucial. For agenda plan-

ning in general, optimizing the agenda as a whole is

more important to the participants, than optimizing the

distinctive journeys between agenda items. Most of

the respondents indicated that inﬂuencing the means

of transportation is not very important to them, as long

as the choice is sensible. A few respondents, how-

ever, expect to have full control over the choice of

transportation. Specifying own time frames for cer-

tain tasks and an enforced order of agenda items were

named as the most important features. Embedding ex-

ternal information, such as a weather-dependent score

was listed as the least important feature.

Flexible Multicriterial Agenda Planning in Public Transit Systems

565

Figure 3 shows the completion time of the created

agenda. It can be seen, that the automatically planned

agenda is slightly better than the average of manually

created agendas. However, one participant of the sur-

vey has created a faster agenda than our implemented

agenda planner, showing that there is still room for

improvement for the agenda planning algorithm. It

shows that, at least when only comparing the comple-

tion time of the agenda, the agenda planner is able to

compute agendas that are slightly better than most of

the manually planned agendas.

Figure 3: Completion time of the manually and automatically

generated agendas.

Figure 4: Planning of the manually and automatically gener-

ated agendas.

Figure 4 shows the time it took for planning the

respective agenda. The agenda planner was a lot faster

than the participants. This comparison, however, is not

representative as the respondents had to use a newly

developed tool, that they were not familiar with. Fur-

thermore, the way the agenda had to be planned was

not as it would be in a real-life setting. Nevertheless,

it shows that planning an agenda is a time-consuming

task, especially in a not so well known location. The

agenda planner took around 5 minutes to compute a

single agenda after 10 non-improving iterations. For

an interactive agent system, these 5 minutes are far

too long. Most of the runtime is spent in assigning

scores to the possible insertions of tasks and in the

computation of journeys between two locations, this

part of the algorithm needs to be improved.

Even though the completion time is similar, people

preferred aspects of their own agenda. One of the most

noted aspect is that the agenda planner planned more

tasks more towards the evening of the day, whereas

most respondents said that they prefer doing their tasks

in the morning. The agenda planner planned the tasks

mostly in the evening, because it was slightly faster to

visit them there. This is due to the ﬁxed location tasks

that had to be visited in the morning and in the evening,

as the user was then already in the vicinity of other suit-

able locations. The manually planned agendas by the

participants were mostly as compact as possible, i. e.

they started the agenda early and ended it early. In con-

trast to that, the agenda planner placed more free time

in the beginning of the agenda and placed more tasks

into the afternoon. Therefore, the continuous free time

was smaller than that of the manually planned agendas,

which was perceived as a negative aspect in the survey.

Since the agenda planner favors agenda items from

the same clusters, the algorithm found clusters where

several of the tasks could be completed, which people

mostly missed. People liked this aspect of the agenda

planner, because this means that the number of pub-

lic transit journeys is reduced and locations which are

reachable with a short amount of walking are preferred.

Another aspect that people have criticized is that there

is not enough room for customizability of the query at

the moment. Overall, the computed agenda was rated

with an average of 4.23 on a scale from 1 to 5, and half

of the people remarked that the computed agenda is

better than their manually planned agenda.

7 CONCLUSION AND FUTURE

WORK

While no expressive conclusion can be drawn from the

evaluation, it shows that planning mobility-oriented

agendas with an intelligent agent is feasible in practice.

The computed agendas were mostly shorter than the

manually planned ones and were rated only slightly

worse by the some of the participants of the survey.

For the future, the evaluation and the algorithm’s per-

formance have to be improved. For this paper, we

have evaluated the mobility agenda planner only with

a single agenda, which only shows that it is able to

compute mobility agendas in this one scenario. There-

fore, a better evaluation method has to be implemented

for our planner. Furthermore, the approach was only

tested for intra-city mobility itineraries.

For future work we have, thus, identiﬁed two main

aspects. The ﬁrst aspect is to improve the evaluation

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

566

of potential solutions to the problem. For this, a user

interface needs to be designed, which allows people to

plan their own itineraries with the agent, so that more

scenarios are covered. In addition to more scenarios,

the review of people’s own agendas will be more use-

ful as people have a stronger opinion about their own

mobility itineraries. Missing functionality and fea-

tures can, therefore, be more easily detected. Research

on the design on such an interface has already been

started in (Wienken et al., 2017). Future work should

evaluate these design guidelines and implement a user

interface for the agenda planner so that people can

interact with it. A user interface, however, may not

be restricted to a graphical user interface. Currently

intelligent personal assistants, such as Amazon Alexa,

Apple’s Siri, and Google Assistant are becoming more

popular. They allow users to interact with software

using natural speech. Such a electronic personal agent

may support the user then in planning their mobility-

agenda. As the personal assistant already has access

to the calendar and the preferences of the user, it al-

ready has much of the required information. These

personal agents already individually support the user

in agenda planning and mobility planning, but an inte-

grated solution that combines these aspects is missing.

The intelligent agent will act more like a recommender

system in such a scenario, because it does not automat-

ically compute the whole agenda, but suggests date

and locations for certain tasks of the user.

The other aspect of future work focuses on the

optimization problem itself. Furthermore, the men-

tioned drawbacks and missing functionalities have to

be addressed in future work, so that people can in-

struct the agent more precisely in how it should plan

the mobility-oriented agenda. The algorithm’s runtime

also has to be improved, as waiting for 5 minutes in an

interactive scenario for the answer of the agent is unre-

alistic. Additionally, the algorithm can be designed to

respect user proﬁles, which we have not regarded at all.

A user proﬁle may either be obtained by explicitly let-

ting the user complete a survey which then constructs

the user proﬁle or by learning the user proﬁle while

she interacts with the agent.

ACKNOWLEDGMENTS

This work was partially funded by the German Fed-

eral Ministry of Transport and Digital Infrastructure

(BMVI) for the project “Digitalisierte Mobilit

at – die

Offene Mobilit

atsplattform” (19E16007B).

REFERENCES

Baena-Toquero, M. J., Muros-Cobos, J. L., Rodr

ıguez-

Valenzuela, S., and Holgado-Terriza, J. A. (2014). To-

wards sustainability in multi-modal urban planners. In

Connected Vehicles and Expo (ICCVE), 2014 Interna-

tional Conference on, pages 492–497. IEEE.

Boussier, S., Feillet, D., and Gendreau, M. (2007). An exact

algorithm for team orienteering problems. 4OR: A

Quarterly Journal of Operations Research, 5(3):211–

230.

Garcia, A., Vansteenwegen, P., Arbelaitz, O., Souffriau, W.,

and Linaza, M. T. (2013). Integrating public transporta-

tion in personalised electronic tourist guides. Comput-

ers & Operations Research, 40(3):758–774.

Gavalas, D., Konstantopoulos, C., Mastakas, K., Pantziou,

G., and Vathis, N. (2015). Heuristics for the time depen-

dent team orienteering problem: Application to tourist

route planning. Computers & Operations Research,

62:36–50.

Golden, B. L., Levy, L., and Vohra, R. (1987). The orien-

teering problem. Naval Research Logistics, 34(3):307–

318.

Gunawan, A., Yuan, Z., and Lau, H. C. (2014). A mathe-

matical model and metaheuristics for time dependent

orienteering problem. In Proceedings of the 10th In-

ternational Conference of the Practice and Theory of

Automated Timetabling, pages 202–217. Research Col-

lection School Of Information Systems.

Li, F., Cheng, D., Hadjieleftheriou, M., Kollios, G., and

Teng, S.-H. (2005). On trip planning queries in spatial

databases. In International Symposium on Spatial and

Temporal Databases, pages 273–290. Springer.

Louren

c¸

o, H. R., Martin, O. C., and Stutzle, T. (2003). It-

erated local search. International series in operations

research and management science, pages 321–354.

Lu, E. H.-C., Chen, H.-S., and Tseng, V. S. (2017). An

efﬁcient framework for multirequest route planning in

urban environments. IEEE Transactions on Intelligent

Transportation Systems, 18(4):869–879.

Sharifzadeh, M., Kolahdouzan, M., and Shahabi, C. (2008).

The optimal sequenced route query. The VLDB Journal

— The International Journal on Very Large Data Bases,

17(4):765–787.

Vansteenwegen, P. and Van Oudheusden, D. (2007). The

mobile tourist guide: an or opportunity. OR insight,

20(3):21–27.

Wienken, T., Kr

omker, H., and Spundﬂasch, S. (2017).

Agenda planning-design guidelines for holistic mobil-

ity planning. In International Conference on Human-

Computer Interaction, pages 713–720. Springer.

Wienken, T., Mayas, C., H

orold, S., and Kr

omker, H. (2014).

Model of mobility oriented agenda planning. In Inter-

national Conference on Human-Computer Interaction,

pages 537–544. Springer.

Zografos, K. G. and Androutsopoulos, K. N. (2008). Al-

gorithms for itinerary planning in multimodal trans-

portation networks. IEEE Transactions on Intelligent

Transportation Systems, 9(1):175–184.

Flexible Multicriterial Agenda Planning in Public Transit Systems

567