Adaptive Assistance for Crisis Response

Wayne Iba and Melinda Gervasio Institute for the Study of Learning and Expertise 2164 Staunton Ct. Palo Alto, CA 94306 [IBA,GERVASIO]~ISLE.ORG

Computer systems are becoming extremely powerful tion and the need to act before complete information as well as more pervasive. In order to maximize their can be collected. benefits, we want to provide adaptive interface layers Naturally, a user interface to a crisis response assis- between these systems and their users. Specifically, we tant should help the user sort through uncertainty, pri- expect the interface layer to adapt to unique charac- oritize actions, and help meet deadlines so as to protect teristics of a particular user. Ultimately, we hope to the threatened value. An adaptive user interface should uncover general principles of adaptive user interface de- modify its assistance so as to accommodateand antic- sign, which include techniques for modeling user habits, ipate a particular user’s strategy, values, and habits. gathering information and feedback to drive those mod- Toward these ends, we have developed INCA, an INter- els, and a methodology for quantifying the value em- active Crisis Assistant, that addresses these concerns. bodied in a particular adaptive user interface. To test INCA, we also developed HAZMAT,a syn- Toward these ends, we have explored several prob- thetic crisis response domain and simulator. The HAZ- lem domains and several modeling techniques. We have MATworld models spills and fires of hazardous mate- tested some of the combinations on users and are in the rials. In this domain, there are 4,000 unique incidents process of designing methods to test the others. Our varying in the type of material involved and the magni- strategy has been to first focus on the hardest problems tude of the spill and fire. The responder has a total of with the assumption that we could learn the most from 49 primitive actions that may be used to contain and these. Thus, we have focused primarily on the crisis control the incident. However, not every action is ap- response domain, a domain where the user (responder) propriate (or allowable) in every incident. Finally, the must cope with threat, uncertainty, and urgency. In user has 25 types of resources available. this short abstract, we sketch our research with respect Responding to a HAZMATincident involves choosing to crisis response, and then close with a summaryof a subset of the actions and scheduling them on the avail- our efforts to generalize our methods and results. able resources without violating any of the resource’s constraints. The effectiveness of such a response is mea- Crisis response sured by HAZMAT’Ssimulator, which compares the con- sequences of the scheduled actions to the outcome when It might be fair to say that crisis response is the most no actions are applied. difficult of problem domains. Organizations and indi- viduals can find themselves in crisis in situations that can be characterized by three elements: threat, uncer- Response assistance tainty, and urgency. First, something of value to an One of our primary concerns is to provide useful as- entity is at risk; unless actions are taken to change the sistance to the crisis responder. Others have utilized course of events, the entity expects the threat to occur. computational planning and scheduling to provide user Second, the entity has considerable uncertainty about assistance. We have focused on attending to the user the details of the situation and the outcomes of possi- and providing only the levels and types of assistance ble actions. Finally, the entity in crisis perceives that if that she finds useful. Our results have shown that in something is not done soon, that the threat will occur; some cases, the computational approach is much worse that is, a sense of urgency or time pressure also charac- than providing less, but "user guided" assistance. terizes a crisis response situation. Therefore, the crisis Our INCAsystem builds three types of knowledge response task is to quickly select a course of action, or about the user. First, it learns about the types of response, that minimizes the expected damage to the complete solutions that a user generates. These are value at risk weighted by the uncertainty in the situa- stored in cases and are used through a case-based retrieval mechanismto initialize future responses. How- Copyright © 2000, AmericanAssociation for Artificial In- ever these initialized responses are typically incomplete telligence (www.aaai.org).All rights reserved. due to minor differences between the situation from the 48 assessment are performed independently. This struc- ResponseInitiation 1 ture capitalizes on the respective strengths of the user and l" and the system, and their synthesis leads to better re- Incident Monitoring[ sponses than either could achieve on their own. / Results summary \ I Plan . . Sohedule/ In numerous experiments involving both human and I Adaptation [ [ Adaptation synthetic users we have established several findings with respect to INCA’Sutility. First, we have found that case- based seeding significantly reduces the time required Case by a user to generate a response at a particular level INCA Retrieva of effectiveness. In the crisis context where urgency is a critical concern, this benefit is especially signifi- state information incidentalarm l and action command cant. Furthermore, we found that a fully automated approach that generated a response and presented it to HazardousMaterials Incident Simulator the user required more time than providing no assis- I l tance at all! It turned out that the solutions generated Figure 1: Mixed-initiative response to HAZMATcrises by the automated system, though equally effective to involving INCAand a human user. user generated responses, were incomprehensible and unsatisfactory to the users. We also found that INCAcorrectly predicted repair actions that were selected by the user. Although we stored case and the new incident. Therefore, the user have not quantified the time savings resulting from this will typically work from the seeded response and make type of assistance, we adopt the premise that recogni- various modifications or repairs. This corresponds to tion is faster and cognitively less taxing than retrieval. INCA’s second type of user modeling. In this case, assenting to the desired repair action when INCAalso learns about the specific types of repairs presented as the default, should be faster than selecting to a response that a user will implement. Internally to the repair action to implement. INCA,a selected repair action is represented by the cur- In both modeling the preferred repair actions and rent state of the response (which actions are scheduled response seeding, we found user-specific patterns. That and resources allocated) and the features describing the is, users tended to benefit more from their personalized incident itself. Based on experience with a user over assistant than from a generic assistant that modeledall numerous HAZMATincidents, INCA forms situation- users. An exception to this finding occurred where the action rules that predict what repair action a given user user was a novice; in this case, we found the novice will initiate in a particular situation. benefited from the models formed from expert users. Finally, INCAforms a model of the user’s value func- Finally, we demonstrated INCA’sability to converge tion to help guide the search for a satisfactory re- on users’ relative value functions over the various out- sponse. Responding to a HAZMATincident is inher- come dimensions of a HAZMATincident. This was es- ently a multi-valued tradeoff. Resources expended to tablished with a number of synthetic users operating contain and control a spill or fire cost money, the ma- with a variety of value functions. Interestingly, INCA terial spilling and possible structures burning have eco- converged on highly accurate evaluation functions even nomic value, and the introduction of hazardous materi- in cases where the user value function could not be rep- als into the ground water or the air through combustion resented in the model space (linear combinations of di- also has a particular (negative) value. In such domains, mensions). there generally is not an agreed upon value function and hence, different users will respond to the same situation Ongoing work in different manners. To form this type of model, INCA We are currently extending our work with INCAin a presents an ordered set of candidate responses to the number of directions. Specific extensions to INCAin- user. The user’s choice of which candidate to implement clude user-guided plan monitoring during execution, in- or improve serves as feedback on the model’s ordering formation gathering actions to address uncertainty, and - either positive whenthe first candidate is selected, or distributed response tasks requiring coordination and negative feedback when another is. Through repeated cooperation among multiple users. observations of a user’s behavior while responding to in- Driven by the desire for general principles of adaptive cidents, INCAforms a model of the user’s value function user interfaces, we have implemented INCAin a second and uses this model to guide the search for a solution. domain and are exploring two other potential domains. Figure 1 shows the organization of INCA and how This exercise has identified aspects that were more do- it relates to the user and domain. Note that plan and main dependent than we anticipated and has pointed schedule adaptation axe done collaboratively by the sys- in several directions where we expect to find generally tem and the user, whereas case retrieval and situation transferable principles and mechanisms.