Centre for Social Services Engineering
Social Services Simulator
Social Services Simulator
The Problem
Policymakers in the area of social service delivery do not have good tools for evaluating the effectiveness of alternative programs before they become policy. Instead, policymakers operate under the assumption that all information required to make a decision is known about impacted service providers and their clients, and that appropriate accommodations can be made for any policy shortcomings. The new policy can then be applied in a targeted location to a broader population of clients than the research and pilot studies it is based on. Also, any clients who respond differently to the new policy act “irrationally” in response to the policy, which cannot be accurately anticipated. In reality, to accommodate these shortcomings, policies targeting social services require a great deal of customization to incorporate the unique circumstances of the targeted social service system, often moving away from the original nature of the policy that was chosen.
There are three main factors preventing the development of tools for evaluating such programs. First, there is insufficient data that captures the unique and dynamic lives of service clients, especially the homeless population. Second, the homeless population is notoriously difficult to make predictions about. Structural factors are often unknown, undocumented, or under-reported. Internal factors for transient populations are often difficult to establish, while negative influences are often underreported. Thirdly, operational models developed by researchers in artificial intelligence, economics, or sociology do not capture the complexity required to model the homeless population. Traditionally, large-scale simulations have been used to try and close the gap between program trials and the implementation of a complete policy. Based on the population’s decisions under past policies, probabilistic models capture a variety of factors that reveal motivations and preferences. Missing factors are supplemented by social science models that may rely on social norms and structural factors more relevant to the general population.
A key difficulty of representing clients is their seemingly “irrational” behaviour in response to well-planned and structured intervention programs. Before a client can be emulated, their behaviour must be understood sufficiently enough to be reproducible. Hence, the focus of this project is to explain this “irrationality” in a way that can be understood through existing models of human behaviour by creating a high-fidelity client emulation model. Towards this goal, the Centre for Social Services Engineering has developed BRAMA, Bounded Rational Agent MotivAtion framework.
Project Description
The BRAMA framework is a high-fidelity simulation to emulate human behaviour. The objective is to emulate a seemingly “irrational” individual with the use of a rational reasoner. The reasoner is an extension of the STRIPS planning algorithm called STRIPS-BR by incorporating four components to emulate seemingly “irrational” human-like behaviour. First, the extension includes a configurable bounded rationality module that captures cognitive limitations exhibited by individuals. Second, Maslow’s hierarchy of needs to categorize and rank agent goals based on five basic human needs. Third, the agent’s planning reasoner relies on a utility function based on emotional changes that an individual may experience while interacting with a social service program. Fourth, to capture the adaptable nature of individuals responding to dynamic and uncertain situations, STRIPS-BR incorporates a replanning algorithm that adapts to cognitive limitations and emotional state of an agent by reranking goals and generating a new plan of behaviour.
Goals
The goal of this project is to develop a high-fidelity homeless person emulation that can be used in a simulation environment to evaluate social service policies. This model must address the under-representation of clients and their needs in existing studies.
Objectives and Results
1. Evaluate applicable decision theories and their limitations. Often people experiencing homelessness are perceived through the filter of social norms by the general population. At the same time, they face different constraints than the rest of the population in their society and live by different social norms. To an outside observer, their life choices may seem irrational, incompatible with society, and detrimental to their own wellbeing. Existing operational models are limited by their unrealistic decision theories when applied to underrepresented populations. Many of the limitations lie in the assumptions that axioms make about either the omniscience of the observer or the independence of individual choices. The theoretical analysis of these axioms is presented. The analysis first focuses on the importance of observable behaviour over assumptions about the internal processing of an individual’s decision making. Second, the analysis highlights how subjective goal and action utility is calculated from a bounded observer’s perspective.
2. Build a framework for emulating seemingly “irrational” behaviour grounded in decision theories.
The premise of this project has been that social service clients, like all individuals, are rational, but are bounded by different limitations and have different beliefs and desires than the general population. The STRIPS- BR reasoner is an extension to traditional STRIPS. STRIPS-BR incorporates all human-centric contributions described above. STRIPS-BR also provides an algorithm to emulate “irrational” behaviour grounded in dynamic subjective and sequential decision theories.
3. Explicitly define limitations based on bounded rationality. Any artificial intelligence system with finite resources is bounded by at least one of the three bounds defined by bounded rationality: time, cognitive limits, and incomplete information. If a domain is small enough or understood well enough, it may be possible to find all possible solutions. However, most problems are not sufficiently well understood, hence most systems must overcome bounds through efficiency improvements. BRAMA is different in that it recognizes client limitations and provides explicit definitions and implementations of these limitations to be used by a rational reasoner. The bounds are not deficiencies of the system but a requirement for a configurable, high-fidelity model. Unlike other systems, the focus is not making human-like reasoning more efficient but making the reasoner itself more human-like. An essential step towards this was the recognition that the observer is not omniscient but bounded like its subjects.
4. Incorporate human-like goal reasoning. In BRAMA, AI planning methods for reasoning about goals are grounded in basic human needs defined by Maslow’s hierarchy. A set of measures were defined to incorporate the hierarchy for goal ordering and utility calculation. A definition of basic needs in the social service domain provided steps for mapping explicit goals expressed by clients to basic needs. This provides an explicit goal ordering. The BRAMA utility function enforces the order by penalizing plans that satisfy goals out of that order.
Three different goal rankings were identified and incorporated into BRAMA. The agent’s preferred goal ranking captures an individual’s unique and subjective preferences, without justification or explanation. Relying on the Calgary Homeless Foundation (CHF) data for clients’ preferred ranking provides much needed empirical grounding of the agent’s ranking and the model’s validation process. Maslow’s ranking provides an objective order of goals, one that could be representative of reality. The mapping of requests made by CHF clients to each MH level is a non-trivial process. Domain-specific mappings were created to associate a client’s request at an appropriate level correctly. Finally, the practical ranking represents the order goals are actually satisfied in once a client moves through the social service system. It takes into account constraints placed on the service providers that impact the scheduling of services and management of resources.
5. Define an explicit relationship between goals and social services from the client’s perspective. To operationalize human-like goal reasoning in the context of social service clients, the nature of that relationship must be captured explicitly with the use of an ontology. Various services offered by social service providers have been captured before. The overall objective of this previous work can be categorized as some combination of modelling, understanding, and optimizing the existing system due to budget cuts or projected growth in volume. Such objectives view the system from the perspective of the service provider. As a result, any study that focuses on client progress assumes service provisioning is working optimally. This may result in insufficient client progress to be prematurely associated with unknown client-specific issues.
The ontology of social service needs (OSSN) provides a view of the service provider from the client’s perspective. Real client needs were identified and mapped to the required services (Figure 1). This approach abstracted away details not visible to clients, focusing instead on the constraints clients face when using services. These may include missing resources, required order of services, and alternative services available to and known by the client. By focusing on client needs rather than available services, different aspects of a client’s needs are identified and explicitly characterized. For example, social needs incorporate the client’s community and identify what needs can be satisfied by their social network. In such cases, an alternative to a service provider should factor in the client’s choices, and the way providers arrange complementary services.
Figure 1: The Ontology of Social Services Needs
6. Calculate utility based on the emotional cycle of change. A BRAMA agent is rational in that it maximizes its utility. What is unique is the incorporation of the emotional cycle of change to calculate that utility. Using a continuous function rather than predefined direct associations between events and emotional responses makes it possible to capture dynamic emotional behaviour over an extended period of time. This was required since a priori emotional behaviour models are not available for individuals living outside of social norms, like the homeless population.
7. Create a human-like response to stressful and unforeseen situations. An issue arises as a result of bounded rationality; a change in behaviour or decision making is required when unforeseen situations arise. To adjust to such situations in a human-like way, BRAMA incorporates a replanning algorithm that uses emotional thresholds (Figure 2). The algorithm allows the reasoner to emulate an agent’s ability to handle stressful situations by considering their emotional state in response to such situations. The emotion-based thresholds add to the fidelity of the model by providing another configurable measure for defining how resilient an agent is to the stresses of executing a plan. The replanning process also controls the reranking of goals during plan execution.
Figure 2: Replanning and Goal Reranking Simulation
Project Team
Recent Publications
Gajderowicz, B., (2019), "Artificial Intelligence Planning Techniques for Emulating Agents with Application in Social Services", PhD Thesis, Technical Report, Centre for Social Services Engineering.
Gajderowicz, B., CA, U., Fox, M. S., & Grüninger, M., (2018), “The Role of Goal Ranking and Mood-Based Utility in Dynamic Replanning Strategies”, Advances in Cognitive Systems, Vol. 9, pp. 211-230.
Gajderowicz, B., Fox, M.S. and Gruninger, M., (2014) "Requirements for an Ontological Foundation for Modeling Social Service Chains", Proceedings of the 2014 Industrial and Systems Engineering Research Conference, Y Guan and H. Liao, eds.