A High Level Symbolic Representation for Behavior Modeling

An important goal of the human behavior modeling community is to develop methods and tools for reducing the expense involved in developing knowledge intensive human behavior models and agent systems. This paper reports an approach addressing this goal. The key feature that sets human behavior models apart from other software systems is the reliance on intensive amounts of knowledge encoded into appropriate structures and processes. We focus on cognitive architectures, because these are the systems that have most emphasized the development of structures and processes that support human-like levels of knowledge and reasoning. Our analyses of existing cognitive architectures demonstrate that they contain many similar components and processes when viewed at an appropriate level of abstraction. We use this level of abstraction to guide our development of a High Level Symbolic Representation for human behavior modeling. HLSR will ultimately provide a high-level knowledge-based language for specifying behavior models and compilers for translating HLSR models into executable code for different cognitive architectures. 1. Problem Statement Building high-fidelity human behavior models is a difficult process. Often, the knowledge that is used to create such models derives from task lists and expert opinion. Normally, a model designer decides on a scenario and then selects and implements potential tasks and strategies that a human would perform in that situation. These tasks delineate the types of procedural and domain knowledge that a model must embody to generate behavior. However, transforming that knowledge to behavior in a computer-executable form is beyond the reach of typical model developers familiar with the subject matter. At this point, model development usually requires significant efforts by a specialist in knowledge engineering techniques. There are many problems with this traditional approach: 1) human behavior models are difficult to create; 2) once created they are often inflexible (i.e., difficult to apply to similar problems or slight mission changes); 3) the knowledge built into any particular representation is difficult to transfer to another representation; and finally, 4) they are difficult for users to customize. This last problem is particularly egregious because the rapid pace of military development combined with the heightened need for information security makes proving new warfare concepts arduously slow. None of these problems are inherent in the development of human behavior models. Rather, they arise because of the complexity of the models, and model developers must be extremely careful if such problems in design and implementation are to be avoided. Our motivating assumption is to pursue approaches that encourage or assist model developers in avoiding such problems in the first place. Our response to these problems is to develop tools and techniques that automate, as much as possible, the translation between high-level model design and implemented behavior systems, thus reducing the required involvement of a knowledge engineer, and improving the flexibility of the defined models. In particular, this report presents our efforts to develop a High Level Symbolic Representation (HLSR). The HLSR provides a formal level of description closer to the level of concepts that a subject-matter expert (SME) can specify. However, because the semantics of HLSR are defined in a specified language, we can then compile HLSR task specifications into executable code within one of several popular computer architectures for specifying human behavior. We are presently focusing on compilation to Soar (Laird, Newell et al. 1987; Newell 1990) and ACT-R (Anderson and Lebiere 1998). Soar and ACT-R are two of the most mature cognitive architectures available, and they provide explicit support for a variety of computational mechanisms and representations that are directly relevant to human behavior modeling. One serious drawback to these architectures is that they provide very low-level programming languages (with primitive operators that are very close to the architecture’s details), that make developing a behavior model a task akin to programming in assembly language. From one perspective, our goal with HLSR is to advance the state of the art in human behavior modeling in a manner similar to how high-level languages advanced software engineering. At one point in time, the idea of creating a formal design-level language that could be automatically compiled into an efficiently running computer program was a significant challenge, but the history of high-level language design has demonstrated that there was really no other choice for the future of software development. We are in a similar situation with current cognitive architectures. Although it remains a challenge to define abstract processes and representations that cut across the requirements for human behavior modeling, the payoffs for development in the long run will be more than worth the effort.