Desiderata for cognitive architectures that are to support the extent of human-level intelligence required in virtual humans imply the need to bridge a range of dichotomies faced by such architectures. The focus here is first on two general approaches to building such bridges — addition and reduction — and then on a pair of general tools – graphical models and piecewise continuous functions — that exploit the second approach towards developing such an architecture. Evaluation is in terms of the architecture’s demonstrated ability and future potential for bridging the dichotomies.

Historically, approaches to human-level intelligence have divided between those emphasizing the mechanisms involved, such as cognitive architectures, and those focusing on the knowledge content, such as ontologies. In this paper we argue that in order to build cognitive systems capable of human-level event-recognition, a comprehensive infrastructure of perceptual and cognitive mechanisms coupled with high-level knowledge representations is required. In particular, our contribution focuses on an integrated modeling framework (the “Cognitive Engine”), where the learning and knowledge retrieval mechanisms of the ACT-R cognitive architecture are combined with integrated semantic resources for the purpose of event interpretation.

Human wayfinding operates very very differently from traditional deterministic algorithms owing to a) restrictions in working memory resulting in subjective regionalized maps, and b)flexible adoption of different navigation strategies. While a number of cognitive strategies have been proposed for human wayfinding, these have been hard to evaluate thoroughly owing to a lack of computational simulation. In this work, we propose a stochastic approach for capturing these aspects, and argue for a memoryless, stationary implementation. In two longitudinal experiments on the same group of subjects, we first estimate the subjective regionalized maps for each subject on the same familiar spatial domain. Later, based on their wayfinding responses, we can estimate the stationary probabilities for different strategies. We apply this algorithm to evaluate three wayfinding strategies proposed in the literature, and repudiate the previously held suggestion that they are followed equiprobably.

This paper discusses the challenge of designing instructable agents that can learn through interaction with a human expert. Learning through instruction is a powerful paradigm for acquiring knowledge because it limits the complexity of the learning task in a variety of ways. To support learning through instruction, the agent must be able to effectively communicate its lack of knowledge to the human, comprehend instructions, and apply them to the ongoing task. Weidentify some problems of concern when designing instructable agents. We propose an agent design that addresses some of these problems. We instantiate this design in the Soar cognitive architecture and analyze its capabilities on a learning task.

For intelligent agents modeled to emulate people, reference resolution is memory management: when processing an object or event – whether it appears in language or in the simulated physical or cognitive experience of the agent – the agent must determine how that object or event correlates with known objects and events, and must store the new memory with semantically explicit links to related prior knowledge. This paper discusses eventualities for memory-based reference resolution and the modeling strategies used in the OntoAgent environment to permit agents to fully and automatically make reference decisions.

An important form of learning involves acquiring skills that let an agent achieve its goals. While there has been considerable work on learning in planning, most approaches have been sensitive to the representation of domain context, which hurts their generality. A learning mechanism that constructs skills effectively across different representations would suggest more robust behavior. In this paper, we present a novel approach to learning hierarchical task networks that acquires conceptual predicates as learning proceeds, making it less dependent on carefully crafted background knowledge. The representation acquisition procedure expands the system's knowledge about the world, and leads to more rapid learning. We show the effectiveness of the approach by comparing it with one that doesnot change domain representation.

The ability to delay intentions and remember them in the proper context is an important ability for general artificial agents. In this paper, we define the functional requirements of an agent capable of fulfilling delayed intentions with its long-term memories. We show that the long-term memories of different cognitive architec- tures share similar functional properties and that these mechanisms can be used to support delayed intentions. Finally, we do a preliminary evaluation of the different memories for fulfilling delayed intentions and show that there are trade-offs between memory types that warrant further research.

One challenge for cognitive architectures is to effectively use different forms of knowledge and learning. We present a case study of Soar agents that play a multiplayer dice game, in which probabilistic reasoning and heuristic symbolic knowledge appear to play a central role. We develop and evaluate a collection of agents that use different combinations of probabilistic decision making, heuristic symbolic reasoning, opponent modeling, and learning. We demonstrate agents that use Soar’s rule learning mechanism (chunking) to convert deliberate reasoning with probabilities into implicit reasoning, and then use reinforcement learning to further tune performance.

This paper argues the position that an essential approach to the advancement of the state of the art in cognitive systems is to focus on systems that deeply integrate knowledge representations, cognitive capabilities, and knowledge content. Integration is the path to aggregating constraints in ways that improve the science of cognitive systems. However, evaluating the role of knowledge among these constraints has largely been ignored, in part because it is difficult to build and evaluate systems that incorporate large amounts of knowledge. We provide suggestions for evaluating such systems and argue that such evaluations will become easier as we come closer to applying usefully new, integrated learning mechanisms that are capable of acquiring large and effective knowledge bases.

One approach to bridging the historic divide between "symbolic" and "subsymbolic" AI is to incorporate a subsymbolic system and a symbolic system into a synergetic integrative cognitive architecture. Here we consider various issues related to incorporating (subsymbolic) compositional spatiotemporal deep learning networks (CSDLNs, a term introduced to denote the category including HTM, DeSTIN and other similar systems) into an integrative cognitive architecture including symbolic aspects. The core conclusion is that for such integration to be meaningful, it must involve dynamic and adaptive linkage and conversion between CSDLN attractors spanning sensory, motor and goal hierarchies, and analogous representations in the remainder of the integrative architecture. We suggest the mechanism of "semantic CSDLNs", which maintain the general structure of CSDLNs but contain more abstract patterns, similar to those represented in more explicitly symbolic AI systems. This notion is made concrete by describing a planned integration of the DeSTIN CSDLN into the OpenCog integrative cognitive system (which includes a probabilistic-logical symbolic component).