Adaptive Decision Support for Planning under Hard and Soft Constraints

We describe the “Welfare to Work” scenario, and the software we are designing to support case managers’ planning for their clients. The Changing World of Welfare to Work President Clinton signed the revised welfare legislation, “Personal Responsibility Work Opportunity Reconciliation Act (PRWORA)” in 1996. This legislation stipulates a set of supports and regulations for welfare recipients that aim to move those recipients into the paid labor force. Through federal block grants to the states, “Welfare to Work” (WtW) recipients may access such services as financial support, health and mental health services, child care, transportation, and literacy and jobskills training. Each recipient, or client, may receive a total of sixty months of services. These 60 months need not be contiguous; clients strategize how to go in and out of the programs to establish economic selfsufficiency and/or maintain later WtW eligibility. The key decision making to access benefits and services occurs during discussions between the WtW clients and program case managers, who play the role of advisors and regulators. The programs are affected by a plethora of frequently changing mandates, laws, rules and regulations. For instance, agencies must maintain a certain proportion of their clients’ time in federally defined “countable” activities, namely those leading directly to employment. These requirements are often at odds with the needs and preferences of the clients. Services recommended by a particular agency or a particular case manager depend Copyright c © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. on local regulations, preferences, and availability of resources and services. The typical case manager has a nearly unmanageable case load, and must thus rely on various shortcuts in advising. These seem to include using broad templates for client needs and ignoring certain data elicited from the clients. Furthermore, case managers have a hard time keeping up with changing regulations and availability of services. Thus, automated decision support tools can make their jobs more manageable. We are working on two components of decisionsupport software: The model-building component and decision support itself. The underlying paradigm for our decision-support component is planning with constraints under uncertainty. Given the 60-month limit on services, it is crucial that advice from case managers or software take into account longer-term plans, rather than immediate gratification. The constraints on planning arise from the client’s preferences and her limitations, and from availability of services. For instance, a low-literacy client cannot be expected to succeed in college courses, nor can a client without a car arrive at a site unreachable by public transportation. Other constraints arise from regulations, such as a limit to how many months of volunteer work a client may use to satisfy the work requirement. Uncertainty arises whenever case managers make judgments about the likelihood of a client’s success in a specific activity. Factors such as dependents’ unstable health, the client’s physical and mental health, and transportation problems affect client participation in advising and success in planned activities. The Welfare to Work world is in constant flux. Legislative and administrative bodies change regulations in response to legal, political and budgetary considerations. Service availability changes from day to day, whether services are sponsored by government agencies or by private organizations and charities. And client preferences shift as clients learn more, and as their family, physical, mental and economic conditions change. We present an overview of current and planned software for decision support for Welfare to Work programs in Kentucky. In Section 2, we briefly describe our approach to building decision-support software for the Kentucky Temporary Assistance Program (Kentucky’s WtW program). We then give a realistic case study and use it to illustrate needed software. Finally, we briefly outline key challenges we are encountering in this process. Planning with Uncertainty and Constraints As mentioned in the introduction, we approach decision support in a Welfare to Work setting, as well as other settings that involve advising as a key component of decision making, as a problem of planning with uncertain information in the presence of constraints. Figure 1 illustrates the architecture of the decisionsupport system for WtW we are currently building. In this diagram, white boxes represent software components, and dark boxes represent data used for inference and planning. There are three major stages in this design: (a) model building, (b) model refinement, otherwise known as knowledge-based model construction or KBMC (Breese 1992) and (c) planning.