Nomination and Prioritization of Goals in a Cognitive Architecture

Goals play an important role in human cognition. Different aspects of human mind influence the generation of goals they pursue, and the goals guide their behavior. In psychology, researchers made significant efforts to study goals and their origin, and cognitive architectures include various facilities to handle goals of artificial agents. One such architecture, ICARUS, supports goal-driven behaviors while maintaining reactivity, and the top-level goals play an important role by guiding the behavior of ICARUS agents. However, the architecture does not cover the origin of its goals or the management of them, and this imposes restrictions like limited autonomy in ICARUS. In this paper, we extend the architecture to provide the capability to manage top-level goals using the notion of long-term, general goals. We show some illustrative examples in an urban driving domain, and discuss related and future work in this direction. Introduction and Motivation Goals play an important role in human cognition. People have ideas on what they want to do or what they should do, and these give rise to many different goals. Such goals, in turn, guide people’s behavior by restricting the space of possible actions to take. Traditionally, psychologists put significant efforts on the study of this process (Simon, 1967; Sloman, 1987; Gray & Braver, 2002 to name a few). As computational frameworks for models of cognition, most cognitive architectures (Newell, 1990), too, have some supports for goals. At the very least, these architectures allow the specification of goals or subgoals that guide the artificial agent’s behavior. But some architectures provide more, including nomination and prioritization of goals. For instance, CLARION (Sun, 2007) has drive and goal mechanisms that correspond to psychological accounts of goal nomination. In Soar (Laird et al., 1986), the top-level operators can act as reactive goals and there are rules that govern their nomination as goals. Another cognitive architecture, ICARUS (Langley & Choi, 2006), operates in a goal-directed fashion, and uses multiple top-level goals. However, the architecture lacks any mechanism to add, delete, or reorder such goals, limiting its capabilities significantly. In this paper, we present the ICARUS architecture with a new goal management mechanism that is reactive to the environment. We extended the existing architectural distinction between long-term knowledge and shortterm structures to goals by introducing general goal descriptions associated with their own relevance conditions. The system instantiates these goals with respect to the current situation of the world and nominates them as its own top-level goals to guide its behavior. The extended architecture also has a new ability to prioritize its nominated top-level goals by modulating their priority values with continuous degrees of match for the relevance conditions. In the subsequent sections, we briefly review the ICARUS architecture and explain the extension for nomination and prioritization of goals in detail. Then we provide some illustrative examples in an urban driving domain. After that, we conclude after a discussion on related and future work. Review of the ICARUS Architecture ICARUS shares its basic features with other cognitive architectures like Soar (Laird et al., 1986) and ACT-R (Anderson, 1993). It makes commitments to its representation of knowledge, memory structures, and mechanisms for inference, execution, and learning. The system provides a computational framework for intelligent agents, which stays constant across different domains. In this section, we review the basic capabilities of the architecture before we continue our discussion on nomination and prioritization of goals. We start with ICARUS’ representation of knowledge and memories that support this, and then cover the architecture’s inference and execution processes. Throughout this section, we show examples from an urban driving domain, which we also use for demonstration purposes in a later section. Representation and Memories The ICARUS architecture distinguishes conceptual and procedural knowledge. Its concepts describe various aspects of the environment, whereas its skills define procedures that are known to achieve corresponding concepts when executed until completion. ICARUS also distinguishes long-term knowledge and short-term structures. Long-term knowledge includes general descriptions of the environment and procedures. The architecture instantiates them and gets short-term structures relevant to the current situation. The distinctions along these two directions result in four main memories in ICARUS. Its long-term conceptual memory stores general definitions of concepts that use variablized objects and their attributes to describe situations. A longterm skill memory houses variablized skills that define general procedures to achieve certain concepts, namely their goals. When the system instantiates these general concepts and skills, it deposits them in the corresponding short-term memories. A short-term conceptual memory stores instantiated concepts, which the system believes to be true in the current situation. A short-term skill memory holds instantiated skills, along with their corresponding goals. For this reason, we often call the short-term memories as the belief memory and the goal memory, respectively.