People

Research activity in 2001

The activities of the Intelligent User Interfaces group concern the definition of techniques for improving the interaction between humans and computers. This activity has been coupled with a design and development activity, aimed at grounding such techniques by applying them to concrete examples.

Four main research topics can be identified: the design of agent-based architectures for the development of complex Web-based systems; the definition of personalization techniques to enhance the interaction with the user; the definition of problem-solving techniques for the resolution of complex problems; the analysis of the dimensions of user modeling.

• The main result concerning agent-based architectures is Seta2000, a general-purpose, agent-based infrastructure for the development of Web-based systems offering complex functionalities, among which the personalization of the interaction with the user. This architecture has been presented in major international conferences, such as IJCAI’01. In Seta2000, wrapping techniques are applied to enable heterogeneous agents to interact with each other in a shared communication language supporting the exchange of synchronous and asynchronous messages. This is essential to support the exploitation and the harmonic coordination, within a Multi-Agent system, of legacy software and of specialized tools for the execution of complex tasks. Moreover, Seta2000 supports the development of goal-based agents, which can display reactive and proactive behavior, therefore enabling the execution of autonomous tasks, in addition to service provision, during the management of the system activities. The goal-based agents are defined by specifying the types of action they can perform and the goals that the various actions achieve. A generic Agent class defines; (i) the structure of the agent state; (ii) the structure of its actions (characterized by goal, preconditions and list of steps to be performed); (iii) an intepreter which, given the current goal to be achieved, selects the action to be performed by the agent. An agent instance is defined by extending the Agent class: this extension consists in the definition of the type of information included in the agents state and the individual action types that the agent can perform. The agent interpreter, inherited from the Agent class, exploits Java threads to speed up the action execution process, therefore supporting a parallel service provision. Moreover, it applies synchronization techniques to suspend its own tasks when their execution is not possible (e.g., because the preconditions of an action are not satisfied). Finally, the interpreter exploits asynchronous communication to trigger the autonomous execution of internal tasks, mainly related to the support of proactive behavior.

• The research in the personalization area has produced interesting results concerning both user modeling and generation of personalized interactions. In particular, the most recent application of user modeling and recommendation techniques has been carried on within the INTRIGUE project, which concerns the design and development of a personalized tourist information server. As far as user modeling is concerned, group recommendation has been addressed, in order to support the recommendation of tourist attractions to possibly heterogeneous tourist groups (e.g., families with young children, groups including disabled people). This type of recommendation is based on the exploitation of sterotypical information about the preferences of typical tourist groups (e.g., disabled have special requirements on transportation systems, etc.) and on the management of structured group models, synthetizing the preferences of the various homogeneous subgroups forming the tourist group. As far as the personalization of the interaction with the user is concerned, personalization rules and dynamic text generation techniques have been applied to produce customized presentations "on the fly", on the basis of the user’s interests and requirements, stored in the group model. In INTRIGUE, such techniques have been applied, in particular, in order to support the explanation of the system’s recommendations to the group. When generating a recommendation list, the system takes into account the preferences of the various subgroups in the tourist group, in order to select the features of the recommended attractions which best fit the (most relevant) group preferences. In this way, the user can identify, for each recommended item, the points of excellence w.r.t. the various subgroups and select the attractions to include in her/his itinerary in an informed way.

The research about user modeling has been developed also within a different project, funded by Telecom Italia Lab and concerning the development of personalized electronic program guides. Within this project, the focus has been placed on the integration of alternative user modeling techniques in order to manage individual user models that rely on extremely different information sources and inference techniques. In particular, we have investigated the integration of stereotypical information about TV viewers (which contributes to the modeling task by predicting typical preferences of the various types of viewer) with dynamic techniques, based on the unobtrusive observation of the user’s zap-stream during the program viewing activity.

• We have investigated the capabilities of problem solving and reasoning within an agent architecture supporting a costumer in a Web store. In particular, we have investigated the issues relevant to configuration of complex products. A complex product can be viewed as a structured entity whose subparts can be complex products in their turn, or atomic ones. Each product that can only be sold "as it is", i.e. for which the customer has no way to change its internal structure, is defined atomic. Thus, eventually, a complex product can be considered as the result of an assembly of atomic ones. Each complex product can be assembled in a huge number of different variants to meet the customer's requirements. Since it is impossible to list all of these variants, an intelligent system supporting the customer should be able to build "on the fly" a description of a complex product meeting the customer requirements (if any). In order to approach this goal, we have developed a conceptual language able to represent both taxonomic relations between the different types of products and partonomic relations beween complex products and each one of its parts. The language offers also the capability of expressing constraints among the components and the subcomponents of the products. Reasoning mechanisms have been developed able to support the user in the interactive configuration: the agent is able to verify the consistency of the requirements on the product specified so far by the user and to configure either the whole product or only that part relevant to the user’s requirements. The reasoning mechanisms exploit both the taxo-partonomic knowledge and the constraints among subcomponents. In order to limit the computational effort, specialized mechanisms have been defined: in particular focusing mechanisms exploiting constraint propagation techniques have been implemented. The configuration agent includes also a graphical interface able to support the user in the incremental definition of the requirements on the product to be configured. Specific reasoning mechanism are able to single out some types of inconsistency between the user’s requirement and the taxo-partonomic knowledge. Additional graphical interfaces have been developed for supporting domain knowledge acquisition.

The investigation of problem solving methodologies has also been addressed in the INTRIGUE project, where an intelligent agenda is being developed to support the user in the scheduling of itineraries. Given the tourist attractions that the user would like to see, the possible constraints concerning her/his visiting preferences, and the opening hours of such attractions, the intelligent agenda proposes itineraries complying with all such constraints. The scheduling of the itineraries is based on the exploitation of constraint-based techniques.

• The goal of the analysis of the dimensions of user modeling is to single out a methodology of the user that need to be modeled, when tackling a specific application problem with specific goals for personalization. Preliminary usability evaluation of small applications (specifically, an active desktop) with limited and simple forms of personalization has been performed, to get a first impression of the techniques that are most promising in these very specific contexts.

During 2000, the research in the above-mentioned areas has been carried on within the following projects:

CAWICOMS (Customer-Adaptive Web Interface for the Configuration of Products and Services with Multiple Suppliers) is a project founded by the European Union within the "Information Society Technology" Programme of the V Framework. It started July 1^st 2000 and will end October 31^st 2002.

The CAWICOMS project involves the following international partners: German Research Center for Artificial Intelligence, GmbH (DFKI), BT Exact, ETIS, ILOG SA, Telecom Italia S.p.A., University of Klagenfurt , University of Torino.

Focus of this project is to enable businesses to market complex customizable products and services by the new ways of electronic commerce. CAWICOMS will develop the technology for:

• A new generation of configurators that will be able to deal simultaneously with multiple suppliers over a network, to interact with other configuration systems as well as with component catalogues, and to deal with different interaction styles and levels, depending on the customer.

• Web-based human computer interaction that will be able to identify the skill-level, preferences, and needs of the customer.

Within the CAWICOMS project, the working unit of the Dipartimento di Informatica of the University of Torino (UTO) focuses its activity on the development of personalization strategies exploited by the CAWICOMS Frontend, in order to customize the interaction with the user of configuration systems.

Several aspects of the interaction can be personalized. For instance, the layout of the interface, the amount of information to be displayed and the type of questions asked during the configuration of the product/service. The first phase of the project has focused on the last aspect, which strongly depends on the user’s knowledge level and interests, and is critical to the usability of the configuration system. The information about the user’s interests and skills, together with the underlying domain-specific knowledge about configuration models, is exploited to predict the user’s choices during the configuration process. The ultimate goal is to reduce the number of questions to the user during the configuration process and to avoide difficult questions which the user might not be able to answer. The decision of the best strategy for the elicitation of information about configuration parameters (e.g., ask the user to choose the preferred value, question her/him about a property of the product/service which is directly related to the parameter and can therefore support the parameter instantiation, exploit a personalized default to set the parameter value automatically, etc.) is made by evaluating alternative elicitation strategies, represented as rules in a rule-based system. As the user interface is dynamically generated during the interaction, the level of detail addressed during the configuration process continuously adapts to the most recent hypotheses about her/his knowledgeability and interests, therefore gaining a reactive adaptation to the user’s needs.

The generation of personalized user interfaces is performed by the CAWICOMS Frontend. The two main tasks performed by the Frontend are:

• the acquisition of the information for the configuration process (by asking the user, or by predicting suitable values, on the basis of the user’s knowledgeability and estimated preferences);

• the presentation of the (possibly partial) solutions produced by the Backend to the user.

Both tasks can be personalized by exploiting specific strategies, given the information stored in the user model. In order to do that, the Frontend has to carry on the following activities:

1. a user modeling activity, aimed to estimate the customer’s interests and knowledge level, on the basis of her/his behavior during the interaction with the system;

2. a conversion activity, aimed to translate (back and forth) the information coming from the Backend (which reasons in terms of configuration problems) into concepts understandable to users, i.e., not strictly related to the management con constraints and configuration problems, but more related to an external (though, possibly very technical) description of the product/service to be configured;

3. a personalization activity, necessary to select the suitable interaction steps, given the problem at hand and the characteristics (expertise, interests) of the individual customer;

4. a dynamic user interface generation activity, as the information presented on the user’s screen must be integrated "on the fly", step by step during the configuration process.

The user inputs and the defaults suggested by the Frontend are handed to the Backend and the distributed problem solving mechanism is initiated. After the calculation of results – that are passed back using the generic data exchange mechanism – these results are also presented to the user in a personalized way, i.e., for instance the technical details are omitted.

The project focused on the re-design of the multi-agent architecture developed within the SETA project (SErvizi Telematici Adattativi, 1997-2000, founded by Telecom Italia) to build adaptive Web based systems.

The basic principles behind the design are:

• the identification of roles needed to offer the main functionalities of the system;

• the definition of agents for these roles;

• managment of the interaction between the agents.

A set of specialized agents have been defined. Each of them has a key role (for instance, the communication with the Web, page generation, etc.) and manages the related activities, using knowledge and technologies specific to the task. In particular the architecture includes:

• an agent managing the interaction with the user from a logical stand point, allowing the system to decide the next step of the interaction based on context of the previous interaction and on the user requests

• a set of specialized agents for the user modeling, the selection of information to present and the Web page generation

• some agents have inference capabilities to decompose and solve complex problems

The activity within the project involved also the definition of the internal architecture of a single agent. This architecture has been designed to enable the interaction of heterogeneous agents and includes a wrapper to provide a unified message protocol between differend kind of agents. The protocol includes syncronous, asyncronous and multicast types of messages. Moreover, the wrapper allows the agents to manage requests in parallel using the Java thread mechanism. Internally the agents can follow different paradigms: action-based or standard Java objects.For example, the User Modeling component needs to provide user data to other agents, but has also some autonomous activities to carry on, such as revising the user model during the interaction with the system: these activities should be managed indipendently from the other agents requests. Other agents, like the Personaliation Agent, responsible for building the personalized Web pages, simply respond to requests from the browser.

A prototype for personalized tourism services on the Web has been developed, based on this architecture. The system presents to the user artistic attractions as well as restaurants or other useful services. The navigation of the tourist site can be done on the basis of different criteria: information can be found geographically or searching by categories (museums, buildings, restaurants, etc.) or combining the two strategies, for example the user can search for all the Baroque buildings in Torino.

The system has reasoning capabilities, since it is able to schedule the visit during a day by taking into account both the user’s choices and other constraints. In particular, the system enables the user to select and insert in the agenda a set of artistic attractions of interest and to ask for a suggestion about the schedule of the day, taking into account the selected items and other constraints the user might have (for example the starting time of the visit, morning or afternoon preferences).

This project has been carried on in cooperation with TI Lab; it started in September 2000 and lasted until August 2001.

The main goal of this project is the definition of personalization techniques for the customization of the future television services and the exploitation of such techniques within a prototype system for the generation of personalized Electronic Program Guides (EPGs). The customization of the EPG will concern the personalized selection of the TV programs to be advertised, on the basis of the user’s interests.

The system is based on a multi-agent architecture, where specialized agents collect data about the available TV programs, monitor the user’s behavior to retrieve information about her/his interests and select the events to be advertised in the personalized EPG, depending on the user’s preferences at the time of day s/he wants to watch TV. The system applies multi-agent technologies for supporting the agent communication and it runs locally to the user’s Set Top Box, where the tasks for the management of the EPGs are executed. The agent-based infrastructure exploited for the integration of the system is JADE, developed at TI Lab.

L. Ardissono, L. Console, I. Torre An adaptive system for the personalized access to news AI COMMUNICATIONS 2001 14(3) 129-147

L. Ardissono, A. Goy, G. Petrone, M. Segnan, L. Console, L. Lesmo, C. Simone, P. Torasso Agent technologies for the development of adaptive Web stores LECTURE NOTES IN ARTIFICIAL INTELLIGENCE 2001 1991 194-213

L. Ardissono, A. Felfernig, G. Friedrich, D. Jannah, R. Schaefer, M. Zanker Intelligent interfaces for distributed Web-based product and service configuration LECTURE NOTES IN COMPUTER SCIENCE 2001 2098 184-188

D. Magro, P.Torasso Supporting product configuration in a virtual store LECTURE NOTES IN ARTIFICIAL INTELLIGENCE 2001 2175 176-188

I. Torre Goals, tasks and Application domains as the guidelines for defining a framework for User modelling LECTURE NOTES IN COMPUTER SCIENCE 2001 2109 260-265

C. Gena Designing TV Viewer Stereotypes for an Electronic Program Guide LECTURE NOTES IN COMPUTER SCIENCE 2001 2109 274-276

Anna Goy Lexical semantics of emotional adjectives Linguistics in Cognitive Science: Proceedings of Student Conference in Linguistics 10 - MIT Working Papers in Linguistics 37 S. Feist, S. Fix, J. Hay and J. Moore Eds. 50-60 MIT Press Cambridge (MA) 2001

L. Ardissono, A. Goy, G. Petrone, M. Segnan Strategie di personalizzazione per sistemi di commercio elettronico sul WEB Simposio HCITALY ’01, Human computer interaction Italy 26-27 Settembre 2001 Firenze Italia SIE Società Italiana di Ergonomia

L. Ardissono, A. Goy, G. Petrone, M. Segnan, P. Torasso INTRIGUE: presentation of tourist information on the Web Atti del Workshop AIIA su Intelligenza Artificiale per i Beni Culturali e le biblioteche digitali 25 Settembre 2001 Bari Italia

L. Ardissono, F. Portis, P. Torasso, F. Bellifemine, A. Chiarotto, A. Difino Architecture of a system for the generation of personalized Electronic Program Guides Proc. of UM2001 workshop on personalization in future TV 13-14 Luglio 2001 Sonthofen Germany

C. Gena, L. Ardissono On the construction of TV viewer stereotypes starting from lifestyles surveys Proc. of UM2001 workshop on personalization in future TV 13-14 Luglio 2001 Sonthofen Germany

D.Magro, P.Torasso Interactive configuration capability in a sale support system: laziness and focusing mechanisms Proc IJCAI 2001 Configuration Workshop August 6, 2001 Seattle USA