Cognition, Simulation, and Cognitive Science

Informatica 29 (2005) 1–yyy 1 Cognition, Simulation, and Cognitive Science Constructing Knowledge and Synthesizing (Cognitive) Models Markus F. Peschl Dept. of Philosophy of Science University of Vienna Sensengasse 8/10 A-1090 Wien Austria Franz-Markus.Peschl@univie.ac.at http://www.univie.ac.at/wissenschaftstheorie/peschl/ Keywords: cognition, cognitive science, empirical research, explanation, explanatory value, knowledge acquisition, method, philosophy of science, simulation, synthetic mode of knowledge acquisition Received: [Enter date] Both the object and the method of cognitive science are identified as features which make cognitive science a highly attractive and fascinating emerging discipline. Both features will be examined more closely: It will be shown that it is by means clear what the object of cognitive science really is. What remains is a quite unsatisfactory answer concerning the effects of cognition and (simulation) models thereof. In the second part the process of theory development in cognitive science will be investigated: it will be shown that structural similarities can be found between the processes of knowledge acquisition in cognitive systems and the processes of theory development in science, more specifically, in cognitive science. Three major modes and strategies of knowledge acquisition (in cognitive systems) will be identified: (i) the empirical, (ii) the constructive, and (iii) the synthetic mode. It turns out that the synthetic mode plays a major role in the process of theory development in cognitive science: there it appears as the method of simulation which is an integral part of cognitive science. An epistemological framework and the implications of the synthetic mode and of the method of simulation will be developed in this paper. Finally, the original question about the object of investigation in cognitive science will be given a more detailed answer. Povzetek: [Click here and Enter short Abstract in Slovene language] 1 The Fascination Science of Cognitive Cognitive Science has received an incredible increase of attention, research funds, and interest over the last years. Comparing cognitive science to classical disciplines, such as cognitive psychology, neuroscience, linguistics, or biology, one has to ask, what is it that makes cognitive science so particularly interesting and attractive. Several reasons can be found: (i) The object of investigation: the process of cognition is (and has always been) of particular interest not only for a small group of scientists, but also for the “simple person”—cognitive science covers phenomena and questions which everybody experiences permanently in his/her own life. “Phenomenon/object oriented approach” towards cognition: the starting point of investigation is—in most cases—not so much an already existing theoretical or methodological framework, but the (cognitive) phenomenon itself. Methods and theories are developed in the course of investigation. Hence, cognitive science follows an object/phenomenon oriented approach rather than a method oriented strategy. (iii) Interdisciplinary approach to cognition: cognitive science is an interdisciplinary undertaking (e.g., Bechtel and Graham 1998, Clark 2001, Thagard 1996, Wilson and Keil 1999, etc.) towards studying the phenomena of life and cognition; this makes it more open for alternative and integrative perspectives considering a wide variety of approaches and conceptual frameworks (both from the natural sciences and the humanities). (iv) Alternative methods for theory development: cognitive science introduces and makes extensive use of alternative means and methods for doing research, conducting experiments, and constructing and developing theories about cognition: cognitive science was among the first disciplines which made simulation a central tool and an intrinsic part of its (ii) 2 Informatica 23 (1999) xxx–yyy M.F. Peschl algorithmic form) which lead to what an external observer would characterize as “intelligent behavior”; e.g., cognition as a connectionist network mechanism or a symbol manipulation mechanism, life as a information transformation process (in a cellular automation), etc. Obviously, we are not satisfied with these answers and not even the most radical representative of the “hard AI” seriously believes that cognition is really symbol manipulation… So, we have to be careful not to confuse cognitive behavior/effects and the mechanisms being responsible for generating these effects with cognition. These mechanisms act as a vehicle for achieving a better understanding offering one possible explanation of how the observed (“intelligent”) behavior can be generated or what the relationship between internal and external parameters is. The aspect of generation of effects, of behaviors, etc. plays a central role in this approach. It is compatible with a constructivist perspective of science, where the generation of a desired behavior of interest (by a mechanism) is part of the process of explanation and “understanding” (cf. Maturana and Varela 1980). In the course of the sections to come it will turn out that the technique of simulation is a key method I this process of generating behavioral patterns. In such a perspective one has to be aware, however, that there has to remain an intellectual dissatisfaction concerning the question what cognition (or life) is. So, in order to understand what cognitive science is actually doing and simulating, we have to proceed pragmatically: it seems that we have to give up these fundamental questions about the “what” of cognition (at least for the moment) and have to “act as if” the computationally generated behavioral dynamics is cognition or life. Under these premises the object of simulation can be identified and described as follows: cognitive models are numerical/logical models having as their object parameters/variables and their relations to each other (in an algorithmic form); these parameters/variables are correlated more or less directly with parameters being the result of empirical investigations having as their object a certain detectable (and quantifiable) aspect of the behavior of a living cognitive system (or a part thereof). research strategy for the process of theory development/verification. This paper will mainly focus on this last issue of alternative methods (i.e., simulation) for theory development in cognitive science and its relation to the process of constructing knowledge in natural cognitive systems. 1.1 Considerations on the Object of Cognitive Science—a Philosophy of Science Perspective Before tackling the problem of simulation and its relation to developing theories about cognition/life, one has to ask the following question first: what is the object of simulation—and, more generally, of investigation—in cognitive science? This question might look naïve at first sight, but it turns out to be substantial for a clear understanding of what we are actually doing in cognitive science. Bechtel et al. (1998, p 3) state that “cognitive science... examines what cognition is, what it does, and how it works.” Hence, in a first approach one could answer our question from above that our object of simulation is cognition (and/or life)… Very soon it becomes clear, however, that it is by no means clear what cognition really is. Is it the process of thinking/living (what is it?), is it the ability to speak, to “understand”, to perceive and manipulate the world, is it the processes taking place in the neural substratum or in a symbol manipulation machine, is it the biochemical mechanisms in the cells,...? It seems that there is quite some confusion about the object of investigation and, thus, of simulation in cognitive science—provocatively speaking: it seems that cognitive science has lost (or perhaps never really found) its object of investigation—hence, it is a question which still remains to be answered (e.g., Bechtel et al. 1998). If it is not really clear what the object of investigation is from a theoretical perspective, let’s try a more pragmatic approach and have a look at the results of cognitive science: what do so-called cognitive models do? In most cases it is the dynamics of variables (or behavioral actions in robotics) being driven by the execution of an algorithm which is interpreted as “intelligent” or “cognitive” behavior; i.e., variables and their values are interpreted as referring to representational/cognitive states, sensory input, and/or behavioral output. In other words, what we are confronted with in simulation models is the effects of cognition rather than with cognition itself (whatever that is). Furthermore, the cognitive model (be it in the symbol manipulation, neural computation, or dynamical systems paradigm) offers an “explanation” in the form of an internal mechanism (being implemented as an algorithm), which is responsible for the generation of these “cognitive effects”. So it seems that we have found an answer for our original question concerning the object of investigation: what we are studying as well as constructing in cognitive science (and Alife) are mechanisms (be it in biological, biochemical, or 2 Acquiring Knowledge About the Environment Now, after having some idea about the object of investigation in cognitive science let´s take a closer look at the methods which are applied for developing theories about cognition. The process of knowledge acquisition (KA)(in general) is the key for achieving a better understanding of the methodological strength of cognitive science, which consists in extensively applying the method of simulation as the main instrument for theory development. In this section it will turn out that a general pattern in the methods for knowledge acquisition can be found both in natural cognitive systems and in (cognitive) science. Hence, we will follow the argumentative strategy to investigate modes of Cognition, Simulation, and Cognitive... knowledge acquisition in cognitive systems in a first step; in a second step these insights will be applied to the field of cognitive science. Before doing that we have to establish an epistemological framework, which gives us an orientation. Informatica 23 (1999) xxx–yyy interaction between the underlying phenomenon/domain ([a]) and the sensory system leads to the impression that there exists an object having certain characteristic features, such as texture, shape, color, smell, etc. These properties are not “real” in the sense of naïve realism; they emerge as a result of this interaction. This does not imply a solipsistic position, however, because the source for these properties are still the “real” phenomena from the hidden domain [a]. In the field of cognitive science the “objects” we are confronted with are “cognitive behaviors” on this level of visible phenomena: e.g., reasoning behavior, learning behavior, intelligent actions, etc. More generally, these behaviors are the effects of cognition and of the non-directly accessible neural domain. As has been discussed in section 1.1, we are always confronted with the effects of cognition and not with “cognition itself” or with the mechanisms leading to these effects—the “visible domain” is exactly the realm where these effects of cognition can be identified. (c) Domain of (primary neural) representations: In this domain the effects of the hidden domain are represented in the primary representation substratum of the cognitive system (i.e., neural representation system). I.e., we are leaving the ontological domain and enter into the realm of knowledge (which is, of course, very primitive on this level). Domain of theories: This is the realm of science where the phenomenon of interest is represented in the form of a theory (“explanans”)—be it intentionally (i.e., inside the cognitive system) or externally (e.g., as linguistic entity, graphical representation, written paper, book, electronic document, etc.). In the context of cognitive science this is the level of cognitive models. They try to provide a mechanism, which is capable of generating the desired/observed phenomena (compare also the perspective on science by Maturana et al. 1980). Their goal is to have explanatory power and predictive value. 2.1 Epistemological Domains in Process of Constructing Developing Knowledge the and In order to achieve a better understanding of what is happening in the process of constructing knowledge it is useful to take a closer look at the relationship between the object of investigation and its theoretical description. In the field of (cognitive) science one has to differentiate between four domains: (a) Hidden domain: In most cases this domain is not directly accessible through our senses and, hence, remains something to be explained: the explanandum. It is the “ontological” domain. This hidden domain can be the underlying microscopic or molecular structure of matter, or even the intention of an author, if the object of investigation is a literary text. This domain has to be assumed, because oftentimes we are confronted with phenomena in our sensible experience [b] which cannot be explained satisfactorily by themselves. In the case of cognitive science this domain can be, for example, the underlying neural mechanisms leading to observable behavior. Domain of the „visible” phenomena: Of course, this domain is not limited to the phenomena which can be perceived by our visual system; any sensory system can perceive a certain aspect of this domain. It is the ontological domain of our everyday perception as it presents itself to our senses. The perceived phenomena are (emergent) effects of the underlying hidden domain/substratum [a] hitting the natural or artificial sensory systems (and by that become “emergent”). In a very strict interpretation this domain does not exist from an ontological perspective: it is only due to the architecture, the functioning, and dynamics of the sensory systems that certain phenomena appear as “having a certain color”, as solid objects, etc. In other words, the (b) (d) 4 Informatica 23 (1999) xxx–yyy M.F. Peschl (d) domain of theories Theory T: s1 -> s2 (c) (neural) r epr esentation repr esentational space human sensory system instrument envir nmental o visible phenomenon (b) visible phenomena/ effects s t s (a) hidden domain hidden dynamics Figure 1: Four epistemological domains. These four domains, which can be identified in the realm of science, have their equivalent in the cognitive domain: domains in science [a] hidden domain domains in cognition non-directly observable process, which cannot be understood; one has the intuition that there is a more complex mechanism or reality “behind” the perceivable phenomenon: this “process/reality behind” is what this domain is about perceivable phenomenon (which one does not fully understand); the entities in this domain are phenomena emerging from the interaction between non-directly accessible processes and the sensory system primary representation of this perceivable phenomenon (e.g., in the sensory systems and in the primary/secondary sensory fields in the cortex) domain of our “knowledge of the world”: contains everyday explanations, rules, assumptions, etc. [b] domain of the „visible” phenomena [c] [d] domain of (primary) representations domain of theories Visible phenomena [b] have to be interpreted as effects of the hidden domain [a] which is not directly accessible for our sensory systems. Almost any scientific activity is based on the assumption that we are not able to directly access the object of investigation (in its deepest nature) and that we are Cognition, Simulation, and Cognitive... only confronted with effects of this object which are generated on a deeper level (e.g., the “visible” phenomenon of light has its cause in electromagnetic activities, which are not directly accessible). It is exactly this “hidden character” which makes this domain [a] so interesting as an object of scientific research. Every scientific as well as cognitive activity aims at constructing possible mechanisms (= theories [d]) which could act as an explanation for this non directly accessible domain. In other words, the goal is to find a theory which (i) fits into the observable (cognitive) dynamics and which (ii) has an explanatory value. Ad (i): the criterion of fitting between the observable dynamics and the theory/model is satisfied, when the model’s predictions are consistent with the empirical findings. Ad (ii): the explanatory value consists in establishing causal relationships between (visible) phenomena which are normally only seen as a—not necessarily clearly causally determined—succession of states over time in [b]. These causal relationships help the scientist to “understand” the observed dynamics by providing a step-by-step explanation of how this dynamics comes about. Now that the epistemological framework has been established as a first orientation, we are going to take a closer look at the processes, which are involved in the acquisition and construction of knowledge (both in cognitive systems and/or in science). Informatica 23 (1999) xxx–yyy For an observer, the question arises: How can the —for the particular system— “relevant” features of this environment be isolated and extracted? How is this knowledge, which is necessary for the organism’s survival acquired and generated? It seems that there is some kind of transformation going on inside the organism that converts those relevant features into knowledge structures, which enable the organism to generate adequate behavior?2 Observing our own cognitive abilities as well as the process of science one can discover that there exists more than one mode of knowledge acquisition. 2.3 Three Modes of Knowledge Acquisition in Cognitive Systems Three different modes of KA can be identified (see also Peschl and Riegler 2001): (i) the empirical mode, (ii) the mode of construction and abstraction, and (iii) the “virtual mode”. 2.3.1 The “Empirical Mode” of Knowledge Acquisition 2.2 Environment and Acquiring Knowledge About the Environment In order to survive, it is a conditio sine qua non for every organism to have some kind of knowledge/ representation1 of its environment. Here is the basic situation one is confronted with when in the process of investigating the representational capabilities of an organism. On the one hand, there is an organism, which has to generate adequate behavior for enabling and ensuring its survival. In order to do so, it is obliged to acquire and make use of knowledge about the structure of its environment. On the other hand there is the surrounding environment having a certain structure and dynamics which is only partially known by this organism. There is a kind of key-lock situation: the organism has to construct knowledge structures (i.e., one possible key) which fit into the environmental structures (i.e., the lock)—it is only by a process of successful unlocking the environmental dynamics that the organism can trigger these environmental processes which are beneficial for its survival. 1 A rather broad notion of the concept of “representation” is used here. It ranges from very simple forms of representation in primitive organisms, such as the embodiment of knowledge in physical or biochemical structures (e.g., the knowledge enabling an organism to follow a chemical gradient), “up to” complex symbolic or logical structures describing the world. The classical mode of knowledge acquisition is what can be referred to as the “empirical approach”. In this mode, a certain aspect of the environment is detected or perceived by a sensory system. In a (natural) cognitive system this process of detection is assumed to be realized in the course of perception by the sensory system and in the first steps of processing these incoming signals. In the context of science, this mode of knowledge acquisition is the most basic and classical approach to any environment: namely, to make an experiment and to use gauges for detecting certain environmental states. What is happening in this mode of KA from an epistemological perspective? It is the very first contact between the environmental dynamics and the representational system (be it natural or artificial). The sensor system is sensitive to a certain state or aspect of the environmental dynamics. Whenever this state occurs in the environment and the sensory system is present at the location and at the moment of this state (transition), a change of states is triggered in the sensory system as well, i.e., the environmental state (transition) is transformed into an “internal signal”; in cognitive systems this process is referred to as the process of transduction. Keep in mind that due to the internal dynamics of the sensory system, its states do not exclusively depend on the environmental states. The dynamics of the sensory system is also determined by its own history of state transitions and, thus, by its current internal state. As an example think of the highly adaptive processes in our visual or auditory sensory systems: Due to external inputs the 2 Of course, this question touches the whole problem of learning in representational systems.We are not going to discuss this problem in detail, however; rather, we will take a more global look at the epistemological modes of knowledge acquisition. 6 Informatica 23 (1999) xxx–yyy M.F. Peschl correlations (within one modality and/or more modalities), trends and patterns (in time and space), etc. Hence, it does not suffice to just collect data by exposing the (natural and/or artificial) sensory systems to the environment, but to process these data/signals in such a way that they provide a basis for generating adequate behavior. Active processes of construction, abstraction, classification, induction, and abduction are necessary for introducing some structure into this unordered set of data. By providing a theoretical context—be it a scientific theoretical framework or already existing common sense knowledge structures—the semantically neutral values/signals are brought into a semantic context. This way, a semantic value or meaning is induced into the data. Furthermore, active processes of (re-)arranging, searching structures and regularities, etc. are applied to these data in order to induce and/or project some (again system-relative) classification, spatial, and/or temporal structure and order into/onto the formerly unordered and uncorrelated set of signals/data. These active processes of construction refer to the huge body of questions concerning the problems of learning, induction, classification, adaptation, evolution, etc.—they will not be discussed here. As opposed to “empirical knowledge”, the constructive mode results in knowledge structures which can be characterized in the following way: (a) The epistemological relationship between the environmental structures and the knowledge is a relationship of functional fitness in the sense of Glasersfeld (1984, 1995). (b) It is not knowledge about a specific entity/state in the environment, but it has “universal/general” character (within a certain domain). Hence, it is knowledge representing some general temporal and spatial regularities or patterns among entities of the primary representation. These regularities are not necessarily “objective” regularities; they have been identified as patterns with respect to the organism’s structures, needs, etc. In this context the constructive as well as the system-relative character of this kind of knowledge becomes especially evident. (c) In general, knowledge does not only capture a static state of the environment, but also describes its temporal dynamics (e.g., in differential equations, in recurrent neural architectures, etc.). This kind of knowledge is about state transitions and patterns of the environmental behavioral dynamics. This makes it especially interesting for tracking and predicting certain phenomena and by that developing an explanatory function. (d) In many cases, this knowledge is not coded in some purely qualitative and/or subjective terms, but is an operational knowledge, i.e., knowledge, internal states of the sensory system changes and becomes less receptive to the input. From an outside perspective this can be interpreted as “adaptive behavior”. From an external observer´s perspective this means that the environmental input is distorted. What is the result of these transduction processes? What kind of knowledge can we expect to be “acquired” from the environment by applying this mode of knowledge acquisition? Both in science and in cognition this process of perception results in some kind of primary representation of the environment (i.e., primary signals with representational function). This means that basic features or states of the environmental structure and dynamics are represented as states in the sensory system and in the primary representational processes. However, from what has been said above follows that these representational states are not some kind of direct mapping of the environmental state—rather they are system-relative (in the sense of theory-laden) states which are modulated by the environmental dynamics (e.g., Riegler, Peschl & von Stein 1999). One can think of this primary representation as a rather unordered collection of data which are (a) referring indirectly, i.e., via the constructive transformation process of measuring or transduction, to certain environmental states. (b) Furthermore, they are coded in the system specific code of the representational system, be it in patterns of neural activations or in electrical charges in a gauge. From an epistemological perspective, there is a fundamental/ontological abyss between the original environmental states and these primary representational states— the process of transduction/measuring bridges this gap between these domains. It is important to note that the environmental signal is distorted already in the very first steps of representation: (a) A complete change of the quality of the environmental signal (i.e., the transformation from the quality of environmental dynamics into the system-relative code of the representational system; (b) the distortion in the process of transduction (e.g., adaptation processes, etc.). In any case, this “primary knowledge” represents the point of departure and the foundation for every other operation in the representational domain. 2.3.2 The “Constructive Mode” of Knowledge Acquisition: Abstraction, Induction, and Construction In most cases the system-relative primary representational states are—per se—rather worthless, because (a) they represent only a certain environmental state at a specific moment in time and at a specific place in space and, therefore, (b) are a relatively unordered and uncorrelated collection of representational states or “data”. What is of interest for generating adequate behavior or a scientific theory, however, is an answer to the question whether a kind of structure is present in these data: classes, Cognition, Simulation, and Cognitive... on which operations can be carried out and which enables us to carry out actions. (e) One of the implications of the operational character of knowledge is that one can make predictions by applying, for instance, mental or mathematical operations on this knowledge. For example, by setting the variables of an equation to concrete values it is possible to derive a concrete result by applying the mathematical operations necessary for solving the equation. This result can be interpreted as a prediction being derived from the general operational knowledge plus the concrete situation (being represented by concrete values in the variables). Predictions are not necessarily restricted to applying deductive or purely logical methods (as an alternative think, for instance, of analogies). (f) From predictions one can go one step further to the manipulation and control of environmental structures and dynamics. I.e., in many cases the general knowledge about the internal mechanisms can be applied in such a way that one can actively control and intervene in the environmental dynamics. Every artist or biochemist who has knowledge about the material he or she is working with is an example for such a behavior. Modern technologies in computer industries or biotechnologies have taken this aspect of control and manipulation of the environmental dynamics to the extreme. The “contemplative character” of knowledge has been traded in for the application and practical dimension of knowledge. It is this “constructive mode” which creates what we usually refer to as “knowledge”—be it a scientific theory or our common sense knowledge about a certain aspect of the environment. Learning, adaptation, classification, and construction are normally these processes, which characterize this “standard mode” of knowledge acquisition. The classical process of theory construction in science follows this constructive mode of knowledge acquisition. It seems that this is the ultimate approach for acquiring knowledge from the environment—there is an alternative mode, however: Informatica 23 (1999) xxx–yyy environment. In this sense, this result can be referred to as being “virtual”. Moreover, our mind, as well as computational methods provide the capacity to explore potential effects of our own (potential) actions without having to physically externalize these (motor) actions in the environment. For instance, whenever we are making plans or whenever the method of simulation is applied in the process of theory construction, we are entering into this new mode of knowledge acquisition. In contrast to the previous modes of knowledge acquisition the environment is completely left aside as a source of new knowledge. The virtual domain becomes an alternative stage for developing new knowledge. Any kind of thought experiment or simulation experiment is an instance of this mode of knowledge acquisition. One could claim that the “constructive mode” of knowledge is located in the virtual domain as well. In a way that is right: The formation of, say, a scientific theory is an operation in the representational space. However, there is always a direct feedback from the environment—it is realized via the verification process of making an experiment. So, what are the new features, which are introduced by this “synthetic mode” of knowledge acquisition (KA)? What is it that makes it a real alternative to the other approaches in the process of KA? There are a couple of answers to these questions: (i) Above all, all processes happen in the domain of virtuality, of the representational space; i.e., in this mode of knowledge acquisition, no further direct interaction with the environment (e.g., via experiments, via behavior, etc.) is necessary; empirical experiments are replaced by “virtual experiments”; externalization of behavior is replaced by thought experiments or simulation experiments: the environmental structure, it dynamics, as well its penetration by an experiment or behavioral action are simulated in the representational space. In the classical approaches, the physical environment plays the role of a constraining factor in theory formation and knowledge acquisition. Whatever knowledge structure functionally fits into these (physical) environmental constraints counts as adequate knowledge or theory. As the physical environment has been “lost” in the synthetic mode of knowledge acquisition, some replacement becomes necessary in order to ensure that the development of knowledge is constrained by an environment-like entity. Hence, there is not only need of an operational and functional model or knowledge of the phenomenon, which should be described (i.e., the result of the constructive approach to knowledge acquisition); above that, a sound model of its environment is necessary as well. Only if this criterion is satisfied, virtual (ii) 2.3.3 The “Synthetic Mode” of Knowledge Acquisition: Generating Knowledge in the Virtual Domain From our everyday experience we know that we can anticipate certain situations or events; i.e., these events take place in our mind exclusively and there is no necessity that they happen (physically) in our environment. As has been mentioned above, prediction is a first step into the direction of that alternative mode of gaining new knowledge: By applying mental or computational operations to our knowledge, theory, model, etc., a result is anticipated and/or predicted which has not yet happened in the 8 Informatica 23 (1999) xxx–yyy experiments become possible: One can explore the effects of one’s actions in the virtual domain. In other words, in a virtual experiment the model/knowledge of the phenomenon we are interested in is confronted with the model of the environment—the representational/virtual domain is the stage of these simulation processes. Hence, the cost of detaching oneself from the physical environment results in the necessity of having to develop a sound model of the environment in which the phenomenon which one is interested in can be found. Of course, this approach has a lot of implications as well as risks from the perspective of philosophy of science and epistemology: e.g., a huge problem of theory-ladenness. Despite these worries, these kinds of models have been applied with increasing success in many disciplines, especially in the fields of cognitive science or ALife. M.F. Peschl 3 Cognitive Science and the Synthetic Mode of Knowledge Acquisition (iii) The new feature which is introduced by this approach is that potential theory/ knowledge spaces can be explored in the virtual space of the representational system (be it in thought experiments of a cognitive system or in simulation experiments, e.g., Peschl 2001). In many cases this reduces the costs as well as the risks drastically, as the direct contact with the environment can be avoided. (iv) The final point concerns the question of how the problem of knowledge construction is approached. Whereas the classical knowledge acquisition processes to follow a rather analytical approach, the mode of knowledge acquisition being discussed here stresses the aspect of synthesis (e.g., Braitenberg 1984). Existing theories, knowledge, theoretical entities, etc. are synthetically rearranged, taken apart, put together, tested in an virtual environment, etc. Many models in the field of cognitive science and ALife (e.g., Langton 1995, Webb 2001) are examples of this synthetic approach. Simulations are implemented in the form of a set of “ingredients” and rules controlling the interactions between them. In the course of the simulation these entities interact with each other and new structures, features, behavioral patterns, etc. emerge as a result of these interactions. What are the implications from these conceptual and epistemological considerations for the field of cognitive science and, more concretely, for the process of theory development in cognitive science? Looking at models in cognitive science and ALife, one discovers that their characteristic and intrinsic property consists in being simulation models; be it a connectionist network or a model from computational neuroscience (e.g., Arbib 1995; Bechtel et al. 2002; Gazzaniga 2000), an AI/expert system (Osherson et al. 1990; Winston 1992; Nilsson 1998), a genetic algorithm (Mitchell et al. 1994), a model of a population of organisms (Langton 1995; Steels 1996), or a dynamical system representing cognitive processes (e.g., Thelen et al. 1994, Port and Gelder 1995, Gelder 1998, Clark 2001), etc. What is the motivation for using and sometimes even preferring simulation models (in cognitive science, ALife, or any other discipline)? What makes them so attractive? Why are they so powerful, influential, and successful in the modern (natural) sciences? From a philosophy of science and epistemology perspective, one of the main reasons for using the method of simulation and for constructing computational models is a lack of full access to a complex phenomenon we are interested in (in particular, if it is the process of cognition). This lack of access can be in the spatial and/or in the temporal domain; above that, this lack of access can be caused by one or more of the following parameters: a lack of theoretical and/or background knowledge, of methods, of resolution in the gauges, a lack of time and funding for complex empirical research, etc. In any case, simulation acts as a means for overcoming one or more of these lacks by transferring parts of empirical research into the virtual domain. Furthermore, from the perspective of philosophy of science simulation does not only open up a new space of possibilities for the construction and verification of theories, but also for a higher explanatory value/potential for the phenomenon to be explained. In the course of the sections to come the advantages as well as problems, questions, and limitations of the simulative approach to theory construction will be examined. Especially in the fields of cognitive science and ALife it becomes evident that simulation has taken a prominent place for developing theories on the basis of computational models, because their object of investigation is—if it is clear at all what it is (see section 1.1)—highly complex, in most cases not very well accessible, as well as very difficult to explain. The Role of Simulation in Cognitive Science In order to achieve a better understanding of the role of simulation in the process of developing theories about cognition, it is necessary to understand how it is embedded in the classical empirical approach in science. To a certain degree, every serious simulation experiment is based on empirical data and theories. Cognition, Simulation, and Cognitive... Informatica 23 (1999) xxx–yyy aspect of reality). Furthermore, it should satisfy the criterion of high explanatory value. One has to be aware, however, that this process, which is characteristic of almost every discipline in the natural sciences, does not bring forth “true knowledge/theories” in the strong sense. This cyclic process is based on the “epistemological tension” between a real phenomenon and its theoretical description. The goal of any scientific endeavor consists in closing this epistemological gap by applying the classical method of conducting experiments in which a theory or hypothesis is tested in interaction with the environmental dynamics (constraining the process of theory development; e.g., Popper 1962; Oreskes et al. 1994). 3.1.1 The Empirical Cycle of Knowledge Construction The lower part of Figure 2 shows the classical (empirical) circle of theory development (compare also the empirical and constructive mode of knowledge acquisition in sections 2.3.1f). In general, the goal of this cyclic process consists in finding an epistemological closure between a certain environmental aspect we are interested in (e.g., a certain cognitive behavior) and its theoretical description. This goal is achieved by mutually adapting the theories under construction and the object of investigation to each other until a kind of consistency between these two domains is reached. This is done by applying the classical means of (empirical) theory construction (cf. section 2.3.2): forming a hypothesis out of previous results and in the context of background knowledge/assumptions, making predictions, and manipulating the reality in an experiment (following the instructions given by a set of methods). That is the moment where the knowledge being represented in a particular theory is externalized: The environmental dynamics is triggered in the experimental situation (e.g., in a learning experiment, in giving a person a cognitive task, in stimulating a certain action in a neuron, etc.). This experimental action causes a change/reaction in which the investigated (aspect of the) environment follows its intrinsic dynamics and exhibits a certain behavior3. By making use of measuring instruments the totality of this behavior is reduced to a particular set of environmental states. These states are detected by gauges and get transformed into numerical/quantitative values. In the following steps these values receive an interpretation in the framework of the original theory. These so-called “results” confirm or falsify the original predictions; statistical and inductive methods are used for the construction of alternative hypotheses or adaptations for the original theory. This cycle represents the classical empirical approach to a particular (cognitive) phenomenon. It is repeated until a fit between the theory and the investigated reality is achieved. It is a feedback process in its nature—it is self-regulating and aims at establishing an (epistemological) equilibrium between the domain of the object of investigation and its theoretical description. This implies that the resulting theories will satisfy only an epistemologically internal criterion4, namely the criterion of functional fitness (e.g., Glasersfeld 1984, 1995; Maturana and Varela 1980; Watzlawick 1984); in other words, this feedback process aims at establishing only an epistemological relationship of fitting/viability between these two domains. This means that a theory is viable, if it provides good predictions (and a means for successfully controlling/manipulating the chosen 3 4 It is exactly this dynamics which is of interest for us as scientists. And not an “epistemologically external criterion”, such as truth. It is “external” in the sense that here exists an external (to this feedback process) perspective which allows for an “objective” determination of the (epistemological) distance between truth (= the particular phenomenon) and the theory describing this phenomenon. 10 Informatica 23 (1999) xxx–yyy M.F. Peschl running simulation simulation results interpretation computer program implementation algorithm inductive/speculative adaptation of model concrete model computational description abstract domain common sense/ background knowledge cognitive theory hypothesis accounting/background theories construction/adaptation of theory interpretation observation gauges prediction application of method experiment Figure 2: The classical circle of developing theories (the empirical approach, lower part), its extension by the method of simulation (upper part), and its interactions. result of the empirical (lower) loop), which gets implemented as an algorithm. The execution of this algorithm/simulation leads to results, which confirm or falsify the (empirical and/or virtual) theory under examination. In turn this might lead to the necessity that the original theory has to be adapted or completely changed. The resulting revised theory acts as an alternative starting point for a new loop of a modified virtual (simulation) and/or empirical experiment. In both cases the criterion of epistemological closure and functional fitness between 3.1.2 Introducing Simulation in the Process of Theory Construction The introduction of the method of simulation implies an extension of the classical feedback loop of theory development: in the upper part of Figure 2 one can see that a kind of deductive-inductive mirror-loop is established, when simulation methods are used for the process of theory construction. Here the starting point is the theory or a hypothesis (in many cases being the Cognition, Simulation, and Cognitive... theory, experimental results, and simulation results has to be satisfied. As can be seen in the models from cognitive science or ALife a fertile cooperation between these two modes of knowledge acquisition and theory development is established (e.g., Gazzaniga 2000). Informatica 23 (1999) xxx–yyy for voltage), what counts on this level of abstraction is only the formal relation between the variables (more or less irrespective of their unit of measurement). This is the level on which simulation models or computational models (of cognition) are based. (iii) Algorithms. In this step the logical functions are transformed into an algorithmic form; e.g., differential equations are resolved into a form which can act as a basis for a computer program. On this level the “referential distance” between the parameters referring to a certain aspect of the original cognitive phenomenon and the variables used in this algorithm is increased one more time. Some of the variables are used only for controlling the internal flow of information processing; only a limited number of variables still refers to a specific aspect of the observed cognitive phenomenon.  Digression 1 One can take this process of abstraction even one step further by transforming all possible configurations of values of variables into a state space. I.e., the system under investigation is transformed into a high dimensional vector space —each point in this space represents a possible configuration of variables referring abstractly to a particular state of the physical/cognitive system. This transformation is well known from the domain of artificial neural networks (ANNs) (e.g., Kolen 1994; Peschl 1997; Hertz et al. 1991) where patterns of activations or weight configurations are transformed into activation spaces and weight spaces, respectively. On this level of abstraction a particular point in the vector space does not have a direct reference to the original neural system under investigation any more. It is a purely formal description of a state, which can have multiple different realizations in the physical domain (be it as a neural system or as any other physical system having a similar structure of states). Digression 2 If these states are connected to each other according to the rules given in the empirical theory (i.e., the introduction of state transitions) and letting aside the metrical dimensions of the vector space, the original system gets transformed into its most abstract form: a (finite) automaton. On this level we are only dealing with states and state transitions; they represent the purely functional dynamics of the original system. In the case of ANNs this can be achieved by connecting all points in activation space with each other according to the rules being implicitly given in a particular configuration of synaptic weights (Kolen 1994; Peschl 1997). Of course, this is a (formally) powerful means of describing the dynamics of a 3.2 Steps Towards a Simulation Model After having examined the embedding of simulation in the overall framework of theory development, we are going to have a closer look at the steps and levels of abstraction which are involved in the course of making a simulation experiment in cognitive science5: (i) Empirical theory. In a first step an empirical theory is constructed by applying the feedback loop having been discussed above. From an epistemological perspective a first abstraction takes place in this process: a certain aspect of the observed phenomenon is reduced to a class of “relevant parameters6”. In other words the quality of the original (physical) cognitive phenomenon or behavior is transformed into a quantitative more or less mathematical or logical description. What makes a theory going beyond a simple description or list of these parameters is that it relates these parameters to each other and, thus, introduces mutual dependencies, regularities, and sometimes a dynamic aspect to the formerly descriptive list of observations. Hence, it provides a description and/or representation of possible mechanisms “explaining” and accounting for the (dynamics of the) observed cognitive phenomena (see also section 2.1). In cognitive science these theories are generated by (cognitive) psychology, neuroscience, etc. Formalization & models. If the theory does not have a mathematical form already its constituents are formalized in this step: this means that the empirically grounded parameters are transformed into variables which are related to each other through mathematical functions and/or logical rules. In this step of abstraction the original quality of the given phenomenon is reduced once more: the set of parameters of a theory is reduced to a set of variables with a specific numerical value or range of possible values: whereas in the empirical theory the original quality of the observed phenomenon was still implicitly present in the form of the unit of measurement (e.g., as “kg” for mass or as “V” (ii)  5 I thank Matthias Scheutz for clarifying discussions on this topic (see also Peschl & Scheutz 2001). 6 It would go too far to discuss the problem of identifying what the “relevant parameters” are for a particular (cognitive) system or process. However, many problems with empirical theories as well as simulation models about cognition/mind have their roots in this reduction of the observed phenomenon to its “relevant” parameters. 12 Informatica 23 (1999) xxx–yyy cognitive system, however, it remains a question what happens to the explanatory value on such a highly abstract level of description. It can be seen easily that the explanatory value of such a description is not very high, as the states and their transitions are only very remotely related/referring to the states of the original cognitive system or behavior. In other words, the physical system’s constraints (in time and space) have become almost irrelevant—functional similarity (on a behavioral level) is the only criterion which counts on this level of abstraction. I.e., whatever mechanism leads to the observed behavior acts as a possible explanation of the cognitive phenomenon under investigation. From this lack of explanatory value it becomes clear why most simulation models are not realized on this level of abstraction. What is interesting in such a description (from a theoretical perspective) is the fact that, if it is possible to find or construct such a description, it (finally and formally) has become evident that the observed system or behavior can be described as a computational process. M.F. Peschl computer program can be interpreted as a virtual experiment. Similarly as in an empirical experiment running a simulation brings forth (simulation) results. These results have to be compared to predictions, expectations, and/or already existing empirical data. If there is a discrepancy, this might indicate that an adaptation or change in the model is necessary. In some cases this might lead to a change in the empirical theory—these changes/adaptations in either domain act as a starting point for a new cycle of experiments. It is exactly this effect of simulation models on the original empirical theory (and, as an implication, on the empirical experiments as well), which makes this method so interesting for the process of theory development. These steps have to be repeated and combined with the empirical approach, until an epistemological equilibrium is reached in the form of a theory fitting in the observed data and dynamics. Furthermore, this theory has to satisfy the criterion of having high explanatory power/value. This is achieved by the necessity of having to transform the given phenomenon into a computational process/description which has its focus on the abstract, structural, and functional states and relationships8 in the observed system/behavior rather than on a huge body of accidental (empirical) details. (iv) Computer programs. The algorithm from step (iii) is transformed into a computer program being coded in a specific programming language. This program gets implemented and compiled on a computer. It is interesting to observe that on this level of abstraction physical constraints begin to play an increasing role (especially constraints being caused by computer technology, e.g., in the speed of computation, limitations in memory space, etc.) whereas in the completely abstract form of an finite automaton neither time nor space are playing a significant role. (v) Running the simulation. Through the execution of the program the dynamics of the model becomes explicit: if the output of the program is combined with a sophisticated graphical display or with a cleverly chosen naming of variables, one has the impression that the observed simulated cognitive behavior is very close to reality (compare simulations in the field of artificial neural networks or artificial life7). Of course, this is just an illusion, as one is confronted only with a well-orchestrated change of values in variables over time being generated by the execution of an algorithm. However, it is this illusion of a dynamic system and its cognitively accessible (graphical) representation, which makes the method of simulation so attractive and useful. 4 Conclusions Three major modes and strategies of knowledge acquisition (in cognitive systems) have been presented: (i) the empirical, (ii) the constructive, and (iii) the synthetic mode. In the course of this paper it has been shown that structural similarities can be found between the processes of knowledge acquisition in cognitive systems and the processes of theory development in science, more specifically, in cognitive science. 4.1 Empirical and Synthetic Mode of Theory Development in Cognitive Science One of the fascinations and characteristics of cognitive science has turned out to be the extensive use of the method of simulation. This goes even so far that a model of cognition is not really accepted by the cognitive science community if it is not based on a simulation model. What is happening in a simulation experiment? Similarly, as in the synthetic mode of knowledge 8 (vi) Simulation results and adaptation of the cognitive model/theory. The execution of the 7 This becomes especially evident, if the output is connected to real motor output, such as in a robot. As has been shown in digressions 1 and 2 in point (iii), one has to be careful not to go too high in the level of abstraction —the purely functional description of a cognitive phenomenon in a finite automaton does not have a very high explanatory value any more (compared to the not so abstract cognitive model, such as a neural network or a symbol manipulation algorithm). Cognition, Simulation, and Cognitive... acquisition (see 2.3.3), the empirical loop is extended/ mirrored into the domain of virtuality and computation. In most cases the phenomenon of interest (in reality) which has to be explained acts as the starting point. If the simulation model is not completely speculative, an empirical investigation and approach to theory construction (see Figure 2, lower part) precedes the implementation of a simulation model. This theory is formalized and transformed into an algorithmic form which gets implemented as a computer program. The (empirically constructed) theory is transformed into a computational model and the empirical experiment is—partly—replaced by a virtual experiment by running a simulation of this model on a computer. The result of this virtual and cyclic simulation process is twofold: (a) It creates predictions for “real world dynamics”. (b) If these predictions are not satisfactory, a possible change in the computational model may be necessary which, in turn, may suggest changes in the original (empirically based) theory. Hence, the results of the simulation experiments have a direct effect on the tested theory itself. In this case, a rewritten version of the theory acts as the starting point for a new cycle of empirical and/or simulation experiments. The execution of this algorithm is what makes simulation such an interesting tool: especially in the context of investigating cognitive/living systems, simulation offers rewarding new insights extending the possibilities of the (empirically based) theory: Whereas the empirical theory describes the dynamics of the system only in a rather static manner, applying the method of simulation unfolds the dynamic aspects of this theory in a very efficient manner. By making such a simulation experiment the implications of the Informatica 23 (1999) xxx–yyy tested theory can be seen without having to conduct experiments on the real object. Hence, simulation models offer the possibility to explore “what-if” questions in a rather inexpensive manner. Concerning the dimension of explanatory power, the method of simulation introduces a real alternative to the classical analytical approach: Instead of acquiring knowledge about a cognitive phenomenon by analyzing it, reducing it, and taking it apart, one tries to achieve understanding by synthesizing the observed behavior by generating it in simulation experiments. 4.2 The Object of Cognitive Science Reconsidered Even though our original question “what is cognition?” and “what is the object of investigation in cognitive science?” has not been answered on a fundamental level, computational/simulation models of cognition allow us to be a bit more specific on these concerns; i.e., to fill the rather abstract “pragmatic definition” of the object of cognitive science (being given at the end of section 1.1) with concrete approaches and models which have been developed over the last decades in the field of cognitive science. It is the method of simulation, which played a crucial role for the development of this diversity of models and theories of cognition. The following table tries to summarize the most important approaches and trends in cognitive science according to the main questions concerning the object of investigation, form of representation, form of dynamics, etc. of cognitive models. 14 Informatica 23 (1999) xxx–yyy connectionist cognitive science and computational neuroscience Symbolic computation: The emergence of global rule-based manipulation of states in a network of symbols simple components (“units”) which are inspired by natural neural systems Through any device that Through local rules in and can manipulate (the changes in the connectivity syntactic form of) symbols between these simple components symbolic structures refer to (i) classical perspective: phenomena in the world in patterns of activations refer a stable and linguistically to the world (“subsymbolic transparent manner representation) (ii) alternative perspective (e.g., Peschl 1997): relationship of functional fitness—neural architecture represents behavioral strategies which are highly adaptive and dynamic  embodiment of knowledge implementation by an extraction of statistical external programmer/ spatio-temporal features of knowledge engineer or user the environment by adaptation processes in the  “borrowed knowledge” (“synaptic”) weight structure Represent the stable truths Develop emergent of the real world properties that yield stable solutions to tasks Traditional/classical cognitive science, GOFAI, PSSH M.F. Peschl dynamic systems approach to cognition What is cognition? What is the object of investigation? How does it work? A history of activity that brings forth change and activity Through the self organizing processes of interconnected sensorimotor subnetworks Dynamics of states and trajectories in the state space embodiment of knowledge How does it represent the world (if at all)? How does the system acquire its knowlege/representation? mutual triggering of structural changes and “resonance” between cognitive system dynamics —environment dynamics Become an active and adaptive part of an ongoing and continually changing world What does a good cognitive system do? Table summarizing the most important approaches, models, and accounts on cognition having been developed over the last decades in the field of cognitive science (see also Varela, Thompson, and Rosch 1991; Thelen and Smith 1994, p 43). References Arbib, M.A. (Ed.) (1995). The handbook of brain theory and neural networks. Cambridge, MA: MIT Press. Bechtel, W. and A. Abrahamsen (2002). Connectionism and the mind. Parallel Processing, dynamics, and evolution in networks (second ed.). Malden, MA; Oxford, UK: Blackwell Publishers. Bechtel, W. and G. Graham (Eds.) (1998). A companion to cognitive science. Oxford: Blackwell Publishers. Braitenberg, V. (1984). Vehicles: experiments in synthetic psychology. Cambridge, MA: MIT Press. Clark, A. (2001). Mindware. An introduction to the philosophy of cognitive science. New York: Oxford University Press. Gazzaniga, M.S. (Ed.) (2000). The new cognitive neurosciences (second ed.). Cambridge, MA: MIT Press. Gelder, T.J.v. (1998). The dynamical hypothesis in cognitive science. Behavioral and Brain Sciences 21, 1—14. Glasersfeld, E.v. (1984). An introduction to radical constructivism. In P. Watzlawick (Ed.), The invented reality, pp. 17—40. New York: Norton. Glasersfeld, E.v. (1995). Radical constructivism: a way of knowing and learning. London: Falmer Press. Langton, C.G. (Ed.) (1995). Artificial Life. An Introduction. Cambridge, MA: MIT Press. Maturana, H.R. and F.J. Varela (Eds.) (1980). Autopoiesis and cognition: the realization of the living. Dordrecht, Boston: Reidel Pub. (1980). Mitchell, M. and S. Forrest (1994). Genetic algorithms and artificial life. Artificial Life 1(3), 267—291. Nilsson, N.J. (1998). Artificial intelligence: a new synthesis. San Mateo, CA: Morgan Kaufman. Oreskes, N., K. Shrader-Frechette, and K. Belitz (1994). Verification, validation, and confirmation of numerical models in the earth sciences. Science 263, 641—646. Cognition, Simulation, and Cognitive... Osherson, D.N. and H. Lasnik (Eds.) (1990). An Invitation to cognitive science. Cambridge, MA: MIT Press. Peschl, M.F. (1997). The Representational Relation Between Environmental Structures and Neural Systems: Autonomy and Environmental Dependency in Neural Knowledge Representation. Nonlinear Dynamics, Psychology, and Life Sciences (NDPSFS) 1(2), 99—121. Peschl, M.F. (2001). Constructivism, cognition, and science. An Investigation of its links and possible shortcomings. Foundations of Science 6(1), 125—161. Peschl, M.F. and A. Riegler (2001). Virtual Science. Virtuality and Knowledge Acquisition in Science and Cognition. In A. Riegler et al. (Eds.), Virtual Reality. Cognitive Foundations, Technological Issues, and Philosophical Implications, pp. 9—32. Frankfurt, New York: Peter Lang. Popper, K.R. (1962). Conjectures and refutations; the growth of scientific knowledge. New York: Basic Books. Port, R. and T.J.v. Gelder (Eds.) (1995). Mind as motion: explorations in the dynamics of cognition. Cambridge, MA: MIT Press. Riegler, A., M.F. Peschl, and A.v. Stein (Eds.) (1999). Understanding representation in the cognitive sciences. New York: Kluwer Academic Publishers. Steels, L. (1996). The origins of intelligence. In Proceedings of the Carlo Erba Foundation Meeting on Artificial Life, Milano. Foundazione Carlo Erba. Thagard, P. (1996). Mind. Introduction to cognitive science. Cambridge, MA: MIT Press. Thelen, E. and L.B. Smith (1994). A dynamic systems approach to the development of cognition and action. Cambridge, MA: MIT Press. Varela, F.J., E. Thompson, and E. Rosch (1991). The embodied mind: cognitive science and human experience. Cambridge, MA: MIT Press. Watzlawick, P. (Ed.) (1984). The invented reality. New York: Norton. Webb, B. (2001). Can robots make good models of biological behavior? Behavioral and Brain Sciences 24(6). Wilson, R.A. and F.C. Keil (Eds.) (1999). The MIT Encyclopedia of the cognitive sciences. Cambridge, MA: MIT Press. Winston, P.H. (1992). Artificial Intelligence (third ed.). Reading, MA: Addison-Wesley. Informatica 23 (1999) xxx–yyy 16 Informatica 23 (1999) xxx–yyy M.F. Peschl