record

Thesis Info

LABS ID
00118
Thesis Title
Situated Representation
Author
Michel van Dartel
2nd Author
3rd Author
Degree
Ph.D. (after public defence on 1 december 2005)
Year
2005
Number of Pages
University
Universiteit Maastricht
Thesis Supervisor
Prof.dr. Eric Postma and Prof.dr. van den Herik
Supervisor e-mail
postma AT cs.unimaas.nl
Other Supervisor(s)
Language(s) of Thesis
English
Department / Discipline
Cognitive Science
Copyright Ownership
Michel van Dartel
Languages Familiar to Author
Dutch (native), English (fluent), and German (conversational)
URL where full thesis can be found
www.cs.unimaas.nl/mf.vandartel/proefschrift.pdf
Keywords
Knowledge representation, situated cognition, agent/robot models, artificial intelligence
Abstract: 200-500 words
The notion of representation is well-defined within the traditional computational theory of mind. However, in the relatively novel theory of situated cognition this is not the case. The focus of this thesis is on the nature of representation in situated systems, i.e., situated representation. In the first chapter the problem concerning situated representation is outlined. The chapter indicates that the cognitive sciences are in need of an operationalisation of the notion of situated representation. To investigate a possible realisation of such an operationalisation the following problem statement is formulated: what is the nature of representation in situated systems? Subsequently, two research questions are formulated to investigate the problem statement: (i) to what extent can we identify where the knowledge resides that is used by a situated system to solve a certain task? and (ii) how is this knowledge accessed and used by a situated system when solving a certain task? Furthermore, in this chapter, the methodology of our investigation is described in terms of five conditions for models of cognition (situatedness, embodiment, cognition, parsimony, and transparency) and a short outline of the three models applied in the thesis (a robot model of active categorical perception, a robot model of the Tower of London task, and a robot model of foraging) is given. The second chapter provides background information. It elaborates on representation issues in situated systems and on the two types of representation that such systems use, internal representation and external representation. This elaboration guides the empirical study of situated representation in artificial systems in the third, fourth, and fifth chapter. In the third chapter, the behaviour of robots in a simple model of active categorical perception is examined. The effective sensorimotor mapping of optimised situated robots clarifies the notion of situated representation. The model provides a unique opportunity to study situated representations, because the perceptual ambiguity in the model forces successful situated robots to represent adequately the information they need to perform a given task. The findings obtained by employing the active categorical perception model and the subsequent analyses lead us to four conclusions: (i) reactive robots can cope with perceptual ambiguity in the context of active categorical perception, (ii) reactive robots can organise their behaviour according to sensory stimuli that are no longer present using the environment as an external memory, (iii) reactive robots incorporating a non-linear sensorimotor mapping are better capable of dealing with perceptual ambiguity in an active categorical perception task than those incorporating a linear mapping, and (iv) sensory state-transition diagrams provide insight into the behavioural strategies employed by reactive robots to deal with perceptual ambiguity and their use of the environment as an external memory. Moreover, the findings obtained by employing the active categorical perception model and the subsequent analysis demonstrate that representation in situated systems can be internal and/or external. This indicates that the operationalisation of the notion of situated representation should allow internal representation and external representation. The fourth chapter studies the nature of internal representation. Internal representation is often associated with planning in symbol manipulation tasks. In order to study the nature of internal representation, in this chapter we investigate representation in a situated robot model of a typical planning task that requires symbol manipulation, the Tower of London task. The results obtained with the situated Tower of London model and the subsequent analyses lead us to conclude that the ability to perform (situated) symbol manipulation by internal simulation of perception and behaviour allows the robot to plan ahead in time. Our second conclusion is that representation of both the current and future states of the environment occurs through the mapping of sensor-array activations to actions. For the current state the activation is received from the environment and for the future state the activation is received from the internally generated expected state. The two conclusions indicate that in order to operationalise situated internal representation, the operationalisation should allow internal simulation of perception and behaviour. The fifth chapter studies the nature of external representation of a situated robot performing a foraging task in a stochastic environment. In order to investigate how the externally represented knowledge is accessed and used by the situated robot, we analyse the robot-environment interaction in the situated model of foraging by two different types of analysis (microscopic and macroscopic). The analyses of the results obtained with the situated model of foraging lead us to three conclusions: (i) macroscopic analysis may predict a universal property that can be explained at the microscopic level by microscopic analysis, (ii) macroscopic analysis may complement microscopic analysis in the study of adaptive behaviour, and (iii) macroscopic analysis may be preferred over microscopic analysis, owing to its power to reveal universal properties. Moreover, the experiment with the situated model of foraging and both the analyses show that external representation may reside in the average properties of the interaction with the environment. Robots in the model of foraging do not represent individual food elements (or their physical locations) by element-specific interaction, but represent the uniform distribution of those food elements in their average interaction with the environment. These findings indicate that in order to incorporate external representation into the operationalisation of the notion of situated representation, the operationalisation should allow representation by the average interaction with the environment. The sixth chapter combines the results of the three investigations reported in the preceding chapters. On the basis of these investigations we formulate a new operationalisation of the notion of representation. The new operationalisation holds that for an entity to be adequately represented by a system, it is implied that the system is able to perform and/or simulate internally the entity-specific interaction with the environment. Four advantages of the new operationalisation over its non-situated counterpart are discussed, these advantages concern: (i) external representation and internal representation, (ii) the representation debate, (iii) situated accounts of cognition and awareness, and (iv) the symbol grounding problem. Thereafter, the two operationalisations are related to each other, from which we arrive at the belief that the operationalisation of non-situated representation should be replaced by the new operationalisation. Furthermore, in this chapter, a discussion on the possible implications of the new operationalisation indicates that the new operationalisation may have implications for the fields of artificial intelligence, cognitive neuroscience, and cognitive psychology. In the seventh chapter we answer the research questions formulated in the first chapter by stating that: (1) we can identify where the knowledge resides that is used by a situated system to solve a certain task to the extent that we can reveal the coordination between the sensory and motor system(s) of a system and relate it to the environmental dynamics, and (2) the knowledge which a situated system uses to solve a certain task is accessed and used by: (i) exploiting the attractors in the interaction with the environment (chapter 3), (ii) simulating interaction with the environment internally (chapter 4), and (iii) exploiting the average properties of the interaction with the environment (chapter 5). Furthermore, in the seventh chapter, we answer the problem statement formulated in the first chapter by stating that in emphasising the role of interaction for cognition in the theory of situated cognition the operationalisation of situated representation is essential. We conclude by stating that, in a situated system, representation is as strongly rooted in the environment as the system itself, i.e., representation is situated in nature.