Heracles II: Conditional Constraint Networks for Interleaved Planning and Information Gathering Jos´ e Luis Ambite, Craig A. Knoblock, Maria Muslea Information Sciences Institute University of Southern California 4676 Admiralty Way, Marina del Rey, CA 90292, USA {ambite, knoblock, mariam}@isi.edu http://www.isi.edu/info-agents/ Steven Minton Fetch Technologies, 2041 Rosecrans Ave., El Segundo, CA 90245, USA Steve.Minton@fetch.com http://www.fetch.com/ Abstract Harnessing the increasing amount of information available through public and private networks to inform decision- making presents a considerable challenge. There is a criti- cal need for a system that integrates and structures diverse information in support of the user tasks and goals. The sys- tem must focus on the relevant information, evaluate trade- offs, and suggest courses of action to the user. Since we cannot ensure that all information and preferences can ever be captured by the system, the planning process should be conducted in a mixed-initiative fashion where the user can explore different alternatives and override the system sug- gestions as needed. In this article we present Heracles II, a constraint-based framework for interactive planning and in- formation gathering. We describe the two contributions that address the limitations of previous work: (1) a hierarchically- partitioned conditional constraint network representation that models the task structure of the application domain, and (2) a constraint propagation algorithm that supports flexible user interaction. Heracles II is fully implemented and has been applied to several practical domains such as travel planning and geospatial information integration. Introduction For any activity there is a wealth of information available through public and private networks. Unfortunately, such information is distributed among many sites, with different data formats, schemas, and semantics. Moreover, informa- tion access per se is of limited value. What is needed is a system that integrates and structures diverse information in support of the user tasks and goals. The system must gather the relevant information, evaluate tradeoffs, and sug- gest courses of action to the user. As an example consider travel planning. There are numer- ous sites with relevant travel information: flight schedules and fares (e.g., www.orbitz.com), hotel locations and rates (e.g., www.itn.com), car rental sites (e.g., www.hertz.com), weather information (e.g., weather.yahoo.com), maps and route planning (e.g., www.mapquest.com), airport parking rates (e.g., www.airwise.com), etc. This information needs be integrated with user preferences, such as preferred air- lines or flying times (e.g., avoid red-eye flights), cost con- straints, and company policies, such as allowable airlines, expense caps, per-diem or mileage reimbursement rates. Al- though the user can visit these sites and take into account all the constraints and preferences manually, it is extremely tedious, error-prone, and time-consuming. A system that queries the remote sites, accesses local information, and en- forces the constraints and preferences is much more desir- able. Planning support and interactivity are the main re- quirements for this kind of systems. In order to support planning, the system must (1) gather and integrate the infor- mation in a coherent structure that captures the tasks needed for the application domain, (2) evaluate tradeoffs and select among different alternative courses action, (3) allow the user to explore and override systems suggestions. To provide flexible interaction, the system must (1) allow the user to input data or change choices at any time during the planning process, and (2) handle information sources which return re- sults asynchronously. In this article, we present an approach to mixed-initiative planning and information gathering based on conditional constraint networks (Mittal & Falkenhainer 1990) that ad- dresses these challenges. The rest of the article is struc- tured as follows. First, we discuss the limitations of closely related previous work, including our own initial approach (Knoblock et al. 2001). Second, we describe the core con- tributions of Heracles II: (1) mapping the hierarchical task structure of the planning domain into a conditional con- straint network, and (2) a constraint propagation algorithm that ensures correct propagation in the presence of cycles, user interactions, and asynchronous sources. Third, we briefly discuss an HeraclesMAPS, and application of Her- acles that integrates geospatial and online data. Finally, we discuss related work, future work, and conclude. Background and Previous Work Our own initial approach to mixed-initiative planning and information gathering was Heracles (Knoblock et al. 2001). Heracles models each piece of information as a variable in a constraint network. The different pieces of information are integrated using constraints. The resulting constraint net- work provides a coherent view of the user activities and cap- tures the relevant information and user preferences. For any non-trivial user activity the number of variables and constraints is very large. Thus, Heracles partitions the network hierarchically corresponding to the task structure of the application domain, in a manner similar to Hierarchi- cal Task Network (HTN) planning (Erol, Hendler, & Nau