An architectural framework for complex cognition
Introduction
According to Sternberg and Ben-Zeev (2001), complex cognition “deals with how people mentally represent and think about information.” Cognition can be complex in a number of ways. In one sense, any task, can be made complex by adding sufficient rules or constraints to the original task description. In these cases, the complexity of the problem arises from the depth of detail and knowledge needed to solve the particular task, and some combination of heuristic and brute force approaches are often sufficient to find the solution. Complexity can also arise as a function of diversity, where instead of a single-task, the agent is faced with a variety of tasks over the course of problem solving. Success in such problems is not based on any single heuristic or algorithm. Instead, there are many studies that show that the choice of representation makes a significant difference (Larkin et al., 1980, Kaplan and Simon, 1990, Elia et al., 2007, Larkin and Simon, 1987, van Someren et al., 1998). Consider any of the spatial reasoning questions commonly found on standardized tests. Solving them is often as easy as creating the right diagram and reading off the answer. Attempting to solve the same question with a logic approach is more complicated and takes more time. This choice of which representation to use is made during the design process by the agent designer, and is very rarely part of the problem solving process of the agent. While this approach is suitable for problems that have fixed tasks, when confronting a diverse set of problems, it is unlikely that the choice of a single representation will produce the correct agent behavior. Thus, dynamically choosing an appropriate representation is a powerful strategy for successfully solving complex cognitive tasks.
Representations can be distinguished at a couple of levels. They can be distinct at the level of representational structure like in the case of logic/predicate calculus representations and array-based spatial representations. In this paper, we refer to this level as the level of the representational formalism. Representations can also be different based on the kinds of operations that are performed on the structure. For example, you can imagine the same predicate calculus representational formalism having a set of logic inference rules for backpropagation or implementing a Bayesian framework. These distinctions are not exclusive though a particular structure often lends itself to a particular kind of processing. Choosing the right representation involves making choices in both these dimensions. However, only the first choice, that of the formalism, is part of the architecture of the agent; the agent architecture must be designed to support the existence of a number of representations that differ at the processing levels by providing the necessary scaffolding (the structural variety). It is this aspect of the representational process that we tackle with the use of the Polyscheme architecture, namely, the existence of different structural formalisms that are then available to the agent to represent and reason about a variety of problem.
The use of appropriate representations also raises a second issue of how results from multiple representations are integrated over the course of the problem solving process. For example, if you have a symbolic representation of the physics of the world (a naive physics model) and a spatial/diagrammatic representation of the layout of the same world, how does the architecture support the integration of reasoning that is done over these multiple representations? Again, there are at least two possibilities. The first way is for the agent to choose the appropriate formalism for each sub-problem (the spatial representation is used to reason about the trajectory of a thrown brick and whether its path intersects a mirror in the world and the symbolic representation for reasoning about the effects of the brick hitting the mirror). Each step of the problem solving process is done by the appropriate representation and integration is simply the modification of a common working memory. The problem with this approach is the original problem of choosing the appropriate representation. The second way, proposed in this paper, is to maintain multiple representations during every step of the problem solving process, instead of selecting any single representation for each step. Integration is more difficult in this case but the agent, as a whole, is more flexible to the demands of the environment and the task. The Polyscheme cognitive architecture provides both – multiple formalisms for representation, and an architectural framework that integrates the results of reasoning at every step to drive the problem solving process forward.
We evaluate the Polyscheme architecture’s suitability for solving a variety of tasks where representational formalism is an important factor in finding successful solutions. In particular, we describe Polyscheme agents problem solving in three tasks from the domains of constraint satisfaction, language understanding and planning problems. These domains were chosen to provide wide coverage over a class of tasks important for both human and artificial cognition. While the individual tasks we describe are fairly straightforward and could be solved using traditional symbolic approaches, our evaluation shows how flexibility in representational formalism can be leveraged by the architecture for these three widely different but important problem domains, and show significant effects even in small-sized problems. The coordination of multiple representations allows for a reduction in inferential complexity, improvement in scaling, and compactness of representation. Further research is directed to show how these benefits extend upward to larger cognitive tasks
Section snippets
Multiple representations
There are a number of dimensions along which two (or more) representations can be differentiated. Consider, for example, representing the layout of a chess board. One common approach is to represent the board as a grid with the rows identified using alphabets and the columns using numbers (or vice versa). Locations are a combination of an alphabet and a number (say A2) and moves are written out as pairs of locations (A2, A4). Another (possibly less elegant) way to represent a chess board would
Polyscheme
The cognitive substrate hypothesis (Cassimatis, 2006) argues that there are a relatively small number of AI problems that if properly solved can be adapted to solve the bigger problem of achieving human-level intelligence. These problems include, but are not limited to, forward inference, reasoning about the physical world, categorization, reasoning about space and time, and the simulation of alternate worlds. The cognitive substrate hypothesis is predicated on two important principles – the
Spatial constraint satisfaction
The objective of the spatial constraint satisfaction problem is to locate a set of objects in a grid space such that they satisfy a set of spatial constraints. While our algorithm is inspired from human techniques in spatial reasoning via the use of diagrammatic/spatial representations, there is no claim to any sort of human modeling of spatial constraint satisfaction. The evaluations were done against more traditional Satisfiability (SAT) approaches, and detailed comparisons and explanations
Conclusion & future work
Complex cognitive tasks often require the right initial representation to find an efficient solution. Finding this representation however is a very difficult task. Our approach to this problem is to use multiple representations at each step of the problem solving process thereby avoiding the need for choosing, a priori, an exclusive representation for the task. Results from each representation are integrated at every step and made available to the entire system. This constant integration allows
Acknowledgements
The research reported in this paper was supported by MURI Grant N000140911029, ONR Grant N000140910094 and AFOSR Grant FA9550-07-1-0072. The authors’ views and conclusions should not be interpreted as representing official policies or endorsements, expressed or implied, of ONR, AFOSR or the Government
References (35)
Intelligence without representation
Artificial Intelligence
(1991)- et al.
Integrating cognition, perception and action through mental simulation in robots
Robotics and Autonomous Systems
(2004) - et al.
The effects of different modes of representation on the solution of one-step additive problems
Learning and Instruction
(2007) - et al.
In search of insight
Cognitive Psychology
(1990) - et al.
Soar: An architecture for general intelligence
Artificial Intelligence
(1987) - et al.
ACT-R: A theory of higher level cognition and its relation to visual attention
Human–Computer Interactions
(1997) - et al.
CVC3
- Barwise, J., & Shimojima, A. (1995). Surrogate reasoning. Cognitive Studies: Bulletin of Japanese Cognitive Science...
A cognitive substrate for human-level intelligence
AI magazine
(2006)- Cassimatis, N. L. (2008). Resolving ambiguous, implicit and non-literal references by jointly reasoning over linguistic...
An architecture for adaptive algorithmic hybrids
Multimodal representations as basis for cognitive architecture: Making perception more central to intelligent behavior
An architecture for problem solving with diagrams
Cited by (11)
A compositional approach to creating architecture frameworks with an application to distributed AI systems
2023, Journal of Systems and SoftwareArguments for the effectiveness of human problem solving
2018, Biologically Inspired Cognitive ArchitecturesCitation Excerpt :Therefore, it is not surprising that an extensive effort has been made to understand the cognitive processes responsible for this ability, and many models of human problem solving have been presented (Hummel & Holyoak, 1997; Ohlsson, 1992; Polya, 1957; Davidson, 1995; Davidson & Sternberg, 1986; Davidson, 2003; Weisberg, 1992; Woods, 2000). Additional models of problem solving and, generally, of human cognitive skills have been proposed by research groups on cognitive architectures like ICARUS (Langley & Rogers, 2005; Langley & Trivedi, 2013; Langley, Laird, & Rogers, 2009; Langley & Choi, 2006), ACT-R (Anderson et al., 2004; Anderson, 1996), SOAR (Laird, 2008; Langley et al., 2009; Nason & Laird, 2005), Polyscheme (Cassimatis, 2006; Kurup, Bignoli, Scally, & Cassimatis, 2011), and others (see Duch, Oentaryo, & Pasquire, 2008; Langley et al., 2009). However, while the current state of research can explain some questions about the scope of problems the human brain can solve, we lack some form of explanations what makes solving (novel) problems so fast, i.e., effective.
An explanatory reasoning framework for embodied agents
2012, Biologically Inspired Cognitive ArchitecturesCitation Excerpt :Looking at this from a robotics point of view, cognitive science provides robots with tools that allow them to better interact with humans. The use of similar representations, for instance, maximizes communicative fidelity because it ensures that one agent can transfer a representation to another agent without losing the relevant relationships that make the representation informative (Kurup, Bignoli, Scally, & Cassimatis, 2011; Kurup, Lebiere, Stentz, & Herbert, 2012). As another example, cognitive robots have been shown to be better partners because they can capitalize on the knowledge that cognitive models of their human counterparts can provide them (Hiatt et al., 2011; Kennedy, Bugajska, Harrison, & Trafton, 2009; Trafton et al., 2006).
Artificial Emotions for Rapid Online Explorative Learning
2021, Proceedings of Machine Learning Research