ROSE – AN INTELLIGENT MOBILE ASSISTANT

Discovering Preferred Events and Finding Comfortable Transportation Links

Bjørn Zenker and Bernd Ludwig

Friedrich-Alexander-University Erlangen-Nrnberg

Haberstrasse 2, 91058 Erlangen, Germany

Keywords:

Mobile navigation system, Route planning.

Abstract:

In this paper, we describe ROSE (Routing Service), an application for mobile phones, which suggests the user

events and locations and guides him to them, using the public transport system. It determines the best possible

transport link and accompanies passengers throughout their entire journey. Further, it reacts in real time to

delays in the public transport system and calculates alternative routes when necessary. For route planning in

this context, we will propose a h

u-optimal algorithm for incorporating non-monotone multi dimensional user

preferences. The algorithm is based on an analysis of theoretical foundations for real world route planning

problems and leads to a new approach of how recommendation and route generation subsystems should be

coupled to increase user satisfaction.

1 INTRODUCTION

State-of-the-art navigation systems provide point-to-

point navigation for their users: given a starting point

and an ordered list of destinations, the navigation

system computes a shortest path between subsequent

points and guides the users from one point to the next

in the list. A major drawback of this type of assis-

tance is that the user has to know where he wants to

go to. Often however, this is not the case. For exam-

ple, tourists in unfamiliar cities cannot apply a navi-

gation system successfully until they have looked up

the places available to visit, i.e. a famous castle. Even

city residents often do not know where to go in order

to spend their leisure time or to settle a private or pro-

fessional matter.

In on-the-move situations without having the

usual sources of information at hand, this is not prac-

tical. To ease the whole process of planning, we

are developing ROSE: it combines recommendation

of events and locations with navigation. While the

provided services can be easily accessed from nearly

every mobile phone, the ROSE server incorporates

and preprocesses data from different web sources, like

live public transport data, event and location directo-

ries and map services.

In many situations, users need to get to more than

a single location. From a theoretical point of view,

this results in a traveling salesman problem. In con-

trast to transport companies, pedestrians are much

more ﬂexible in planning their tour: They do not have

ﬁxed time tables, so the problem of satisfying many

constraints can be relaxed: the ordering of visiting the

locations is arbitrary, so tour plans may be changed at

travel time. Often, even the location is not ﬁxed be-

cause users want to satisfy “tasks” (like buying printer

paper or getting something to eat) without having a

certain shop in mind. This allows tour plans to be

rearranged interactively: at any time, the system can

propose a near-by shop for printer paper and settle one

item in the total list of matters.

In this paper, we present an iterative algorithm

that approximates graph theoretical optmization prob-

lems of tour planning by applying the relaxations just

described. Before discussing the algorithm, the pa-

per presents the ROSE system and its current state of

implementation comparing it to other state-of-the-art

systems. Next, we discuss differences between op-

tizimation criteria that are user adaptive and admissi-

ble heuristics that are used regularly for ﬁnding opti-

mal solutions in graphs. We give an account of the

underlying hierarchy of graph-theoretical problems

and outline how the idea of relaxing the optimization

might help in developing tractable algorithms. We

explain our implementation of optimzing user adap-

tive criteria as a multi-attribute decision problem and

present examples encoded in our programming lan-

guage MADL for multi-attribute optimizations.

365

Zenker B. and Ludwig B. (2010).

ROSE – AN INTELLIGENT MOBILE ASSISTANT - Discover ing Preferred Events and Finding Comfortable Transportation Links.

In Proceedings of the 2nd International Conference on Agents and Artiﬁcial Intelligence - Artiﬁcial Intelligence, pages 365-370

DOI: 10.5220/0002735603650370

 SciTePress

Figure 1: Two routes computed by Google Maps: On the left Google’s best recommendation; on the right Google’s second

best recommendation.

2 RELATED WORK

In this section, we argue that the issues of user adap-

tive heuristics for planning in assistance systems and

combination of services from a user-oriented perspec-

tive have not been considered extensively enough in

research on intelligent assistance systems.

2.1 The Need for User Preferences in

Multi-attribute optimization is a well established area

of research. However, as efﬁcient search algorithms

need admissible optimization function for heuristic

evaluation of partial solutions in order to efﬁcient

prune the search space, weighted sums are the pre-

ferred type of heuristics. They provide a clear ad-

vantage: as they are a mapping from a set of inﬂu-

ence factors to the real numbers, a set of options the

search algorithm has to decide between can always be

ordered totally whereas multi-attribute decisions just

compute a pareto-optimal solution. In general, it con-

tains more than a single element. Therefore it remains

unclear which decision is the best for optimal pruning.

On the other hand however, weighted sums suffer

from a important disadvantage: they are bad at ex-

plaining their decisions and – even worse – at allow-

ing the user to take inﬂuence on a decision. To illus-

trate this point, let us look at an example taken from

the Google map server (http://maps.google.de):

we are asking for a route (using a car) from be-

tween the three German cities Weienburg, Winds-

bach, and Herzogenaurach. The route considered best

by Google can be seen on the left map in Fig. 1. It is

123 km long and takes one hour and 42 minutes for

a normal car. The second best recommendation is the

map on the right side in Fig. 1. It is 90 km long and

takes one hour and 45 minutes: the advantage of 3

minutes is reached at the cost of 33 km! This exam-

ple is a case for multi-attribute decisions which allow

the user to state explicitly what his preferences are.

Depending on the driver, highways may be preferred

(taking long distances into account) or not (preferring

to save fuel or just disliking the high speed on German

highways). Generalizing this example, in Sect. 4 we

present an algorithm that incorporates user-adaptive

preferences into heuristics for pruning search spaces.

2.2 Comparison to Similiar Systems

ROSE’s application domain is also treated by other

researchers. Similiar aproaches for recommenda-

tion and navigation systems have been implemented

in COMPASS (van Setten et al., 2004) and Magitti

(Bellotti, 2008). They employ different prediction

strategies in recommendation and rate the yielded re-

sults based on user preferences. P-TOUR (Maruyama

et al., 2004) uses a genetic algorithm to ﬁnd dif-

ferent near-best solutions and presents them to the

user in a k-means clustered overview, from which he

can select his preferred route. RouteCheckr (Voelkel

and Weber, 2008) employs a multi-criteria Dijkstra

based algorithm, which remains limited to the usage

of the weighted sum. As it does not use an estima-

tion function, it does not have to cope with nonmono-

tone and optimal criteria, but is presumably slower

than an algorithm using such a heuristic. Hochmair

(Hochmair, 2004) studies, which decision rules bicy-

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

366

clists utilize in route planning and discusses different

decision rules. He concludes that a compensatory de-

cision rule should be used, but he does not implement

this concept in a new algorithm. These systems do not

support public transport.

In contrast, PECITAS (Tumas and Ricci, 2009)

is a mobile personalisable navigation system, which

advices routes using means of public transportation.

However, it does not include recommendation of

events or locations and routes are restricted to one

starting point and one destination point only. For user

adaptation, PECITAS generates multiple routes by us-

ing different heuristics (e.g. fastest route, or not tak-

ing any bus) and ranks them according to user prefer-

ences (walking preferences, number of bus changes,

arrival at destination, sightseeing).

Compared to these systems, the unique features of

ROSE are:

• routing to multiple destinations,

• integration of recommendation, route generation

with live public transport

• support and navigation

• usage of various compensatory decision rules.

3 OVERVIEW OF THE ROSE

SYSTEM

In this section, we present a short overview of the cur-

rent implementation of ROSE and its client-server ar-

chitecture. In a sample session we demonstrate how

ROSE is used in practice.

3.1 Pedestrian Navigation, Event

Recommendation, and Live Public

Transport Routing – All in One

To get a recommendation, the user enters a query, like

’modern opera’, into his mobile phone (see 2, left).

The recommender then generates a list of suggestions

based on the user input and the user’s preferences. In

this example it would likely be a list of events which

feature music similar to the users preference (see 2,

right). After the user has chosen one of the presented

options, the system calculates a route from the cur-

rent location to the selected goal. Figure (see 3, right)

shows a route overview from the users current point

to his choosen destination. To consider user prefer-

ences in route generation, we propose a h

u-optimal

algorithm in section 4.

To ease the travelling, public transportation is also

considered. The system calculates a route from the

user’s current position to the nearest public transport

option, which means of transportation to take, where

to change transportation and how to walk from the

last stop to the goal location. Departure times are dis-

played to the user and he is informed, i.e. if he has to

hurry to catch a bus.

3.2 System Architecture

The ROSE system consists of a server, which calcu-

lates recommendations and routes and a client which

is used for user input, display of results and naviga-

tion. More details can be found in (Ludwig et al.,

2009).

3.3 The Core Issue: User Adaptive

Optimization

We structured the system into three main services:

recommendation, route generation and navigation. In

the current release of the system, all services are cou-

pled loosely: the results of the recommendation are

the input (goals) of the route generation. The result of

the route generation is the input (way) of the naviga-

tion service. Such a loose combination suffers from a

number of drawbacks:

• It does not provide for replanning when some un-

foreseen event (e.g. missing a bus, the printer pa-

per shop being closed exceptionally) happens.

• The algorithm is challenged by a huge search

space: the fact that busses travel according to

time tables results in a search graph with a quite

tremendous number of edges.

• In most cases, there are many locations meeting

the user’s preferences. For example, many restau-

rants sell good pizza. As a consequence, even the

problem an optimal solution is searched for is not

deﬁned uniquely.

• Finally, the selected locations may be far away

from each other, allowing for many degrees of

freedom in ordering them to form a tour.

Our conclusion from all these issues is that a close

coupling of recommendation is needed which inter-

leaves route generation and navigation. This results

in a complex graph theoretical optimization problem

like the Multi Path Orienteering Problem with Time

Windows (MPOPTW) (Garcia et al., 2009). (Ludwig

et al., 2009) presents an hierarchical overview (com-

pare Figure 4.) of different problem classes in the

routing context.

Additionally, the impact of close and loose cou-

pling of the recommendation and route ﬁnding ser-

vices on the user satisfaction shall be researched. As

ROSE - AN INTELLIGENT MOBILE ASSISTANT - Discovering Preferred Events and Finding Comfortable

Transportation Links

367

Figure 2: left: Query, right: List of recommendations.

Figure 3: left: Details of recommendation, right: Route overview.

we presume that the algorithmic complexity of close

coupling is too big and that loose coupling is not sat-

isfactory, we want to investigate, how an intermediate

coupling between these the extremes could be real-

ized.

4 ALGORITHMS FOR

COMFORTABLE ROUTES

Criteria for evaluating the quality of a route are lim-

ited mostly by formal constraints dictated by the al-

gorithm used to ﬁnd optimal paths. Efﬁcient greedy

graph search algorithms require the heuristic function

to be monotone; A* even requires the heuristics to be

optimistic, i.e. to never overestimate real costs of a

path. In practice however, such constraints for heuris-

tics are not adequate to reason about user preferences.

In a survey conducted at our computer science insti-

tute among public transport users, the following crite-

ria were marked as important by the test candidates:

• No long waiting time until departure

• Short duration of the trip

• Short foot walks

• Few changes of transportation

• No long waiting time during changes

Optimistic estimates for these criteria are just the

function f (x) = 0; this amounts to omitting the crite-

rion completely, which is an undesired consequence.

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

368

Figure 4: Routing Problems for Different Needs.

A second important ﬁnding of our study is that

users do not evaluate the utility of a route on a one-

dimensional scale (where there is always an opti-

mum), but try to ﬁnd a compromise between mul-

tiple attributes that are not comparable among each

other. They accept locally sub-optimal proposals op-

timizing the beneﬁts of a proposal and minimizing its

risk globally.

For example, somebody traveling with a lot of lug-

gage accepts using a bus line that arrives some min-

utes later at the train station than the fastest one, but

is much less crowded. Obviously, in this context, the

comfort of the trip is valued higher than the dura-

tion. From an algorithmic point of view, this means

that the standard shortest path approach cannot be ap-

plied successfully in order to satisfy the user needs

as good as possible. However, searching according to

a heuristic function that forces the search procedure

to visit (almost) the whole search space is no attrac-

tive option for developing programs intended to run

in real-time.

4.1 An h

u-optimal Algorithm for

Comfortable Routes

The key to an efﬁcient solution that belongs to the

complexity class as A* and retains it’s soundness

and completeness is therefore to use a) a monotone

heuristic function h for computing correct solutions

and b) to incorporate a non-monotone, multi-attribute

heuristic u for the user preferences into the search pro-

cedure.

In Fig. 5 you see two paths from the start node s

to a goal node g. g

denotes the actual costs from s

to the node which is currently being expanded, h

the

estimated costs to g according to the heuristic func-

tion h. u

denotes the actual valuation from s to the

node which is currently being expanded, t

the esti-

Figure 5: h

u-optimal Expansion of the Search Space.

mated valuation to g according to the multi-attribute

heuristic function u (user preferences).

If there is exactly one optimal successor state with

+ h

− g

− h

| > ε, the search procedure works

as usual. In the case of |g

+ h

− g

− h

| ≤ ε, both

paths from s to g in Fig. 5 are undistinguishable in h.

In practical applications, if |g

+ h

− g

− h

| ≤ ε

then often other criteria become relevant for deci-

sions. In order to explain the meaning of ε, let us

consider the two public transport connections from s

to g in Fig. 5: one takes 50 minutes, the other one 47

minutes. However, the ﬁrst connection requires fewer

changes of transportation than the second one. If the

user dislikes changes, the ﬁrst connnection is optimal

according to the user preferences u under the assump-

tion of an ε-tolerance ε ≥ 3minutes.

In order to select a path that is both optimal in the

sense of h given ε and in the sense of u, we compute

the ranking p

= t

+ u

and p

= t

+ u

of both op-

tions. The vectors p

and p

represent all user prefer-

ences in u. Each dimension represents a single prefer-

ence (cf. the list of criteria in section 4). As the pref-

erence space to which p

and p

belong is partially or-

dered, in general there is no unique minimal element.

Therefore, we compute a pareto-optimal set which in-

cludes all paths which cannot be distinguished neither

by h nor by u. The ﬁnal selection of a path which is

-optimal and u-optimal is performed by applying a

decision procedure. For calculating decision proce-

dures, we use the MADL programming language, as

described in (Ludwig et al., 2009). We call the result

u-optimal.

4.2 An Approximation of the

u-optimal Algorithm

Instead of implementing the h

u-optimal algorithm

directly, we simulated it’s behavior by having the path

search procedure compute the n-best list of solutions

for a given destination. Each of these n solutions is

ROSE - AN INTELLIGENT MOBILE ASSISTANT - Discovering Preferred Events and Finding Comfortable

Transportation Links

369

evaluated according to the user preferences. The ﬁnal

set of m < n solutions is the set of m routes which are

u-optimal.

In out ﬁrst prototype, we implemented the n-best

approach in order to obtain an evaluation platform as

fast as possible. The algorithm works in two steps:

1. Compute n Best Results

Just computing the n best routes using always the

same heuristics often leads to proposals that only

differ minimally among each other — in particu-

lar, if just one line serves as public transport to the

destination. Therefore, it is better to compute n

routes using n different (optimistic) heuristics and

to compare the n resulting best routes. This obser-

vation has also been made by (Tumas and Ricci,

2009).

2. Rank the n Best Lists

In order to get a global score for each user prefer-

ence and each route, we sum up the contributions

of each segment of the route to each criterion. The

sum is called the rating of the route correspond-

ing to the criterion under investigation. Finally, a

total score is computed: The easiest approach to

take multiple criteria into account is to compute a

weighted sum of all criteria by multiplying each

rating with the weight for the preference as en-

tered by the user in the conﬁguration dialog for

the ROSE system (see Figure 1 left).

A more elaborate approach is to compute a pareto-

optimum for the rating of the n routes. Beyond

that, it is possible in our system to apply other

multi-attribute decision rules, as the Take The Best

decision rule proposed by (Gigerenzer and Gold-

stein, 1996).

5 CONCLUSIONS AND

OUTLOOK

ROSE is a recommendation, route planning and nav-

igation system for mobile devices. To integrate user

multiple preferences and decision rules into the route

planning process, we developed a h

u-optimal algo-

rithm.

We tested ROSE on a major event with 20.000

people to recommend subevents and calculate routes

using the local bus network. We conducted a user

study, from which we hope to get more information

about the relevance of different user preferences in

recommendation and navigation.

REFERENCES

Bellotti, V. (2008). Activity-based serendipitous recom-

mendations with the magitti mobile leisure guide. CHI

2008 Proceedings, 21:498–516.

Garcia, A., Arbelaitz, O., Otaeguui, O., Vansteenwegen, P.,

and Linaza, M. (2009). Public transportation algo-

rithm for an intelligent routing system. 16th ITS world

congress, September 21-25, Stockholm, Sweden.

Gigerenzer, G. and Goldstein, D. (1996). Reasoning the

fast and frugal way: Models of bounded rationality.

Psychological Review, 103:650–669.

Hochmair, H. (2004). Decision support for bicycle route

planning in urban environments. Proceedings of the

7th AGILE Conference on Geographic Information

Science, Crete University Press, Heraklion, Greece,

pages 697–706.

Ludwig, B., Zenker, B., and Schrader, J. (2009). Recom-

mendation of personalized routes with public trans-

port connections.

Maruyama, A., Shibata, N., Murata, Y., and Yasumoto, K.

(2004). P-tour: A personal navigation system for

tourism. Proc. of 11th World Congress on ITS, pages

18–21.

Tumas, G. and Ricci, F. (2009). Personalized mobile city

transport advisory system. ENTER Conference 2009.

van Setten, M., Pokraev, S., and Koolwaaij, J. (2004).

Context-aware recommendations in the mobile tourist

application compass. Adaptive Hypermedia and

Adaptive Web-Based Systems, pages 235–244.

Voelkel, T. and Weber, G. (2008). Routecheckr: personal-

ized multicriteria routing for mobility impaired pedes-

trians. Proceedings of the 10th international ACM

SIGACCESS conference on Computers and accessi-

bility, pages 185–192.

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

370