========================================================================== SUMMARY OF WEBB, B. (2001) "CAN ROBOTS MAKE GOOD MODELS OF BIOLOGICAL BEHAVIOUR?" ========================================================================== CITATION -------- Webb, B. (2001) Can robots make good models of biological behaviour? Behavioural and Brain Sciences 24, pp.1033-1050, 2001, Cambridge University Press. WHAT USE ARE MODELS ------------------- Models help us deal in various ways with a system of interest. How? 1) convenient/manegeable/cheap... 2) but why are they relevant to the target system? Modelling aims to make the process of producing preditions from hyptheses more effective by enlisting the aid of an analogical mechanism. Generally we can more easily manipulate mathematics than physical or chemical models so results of a particular configuration can be more easily predicted. Giere: The hypothesis is a model of which the target system is a type. The process of prediction is then described as "demonstration" of how this hypothetical system should behave like the target. This process of demostration might involve the necesity of a second level of representation, also called a model, in our case a robot. "Simulation is intended to augment our ability to deduce consequences from the assumptions expressed in the hypothesis: "a simulation program is ultimately only a high speed generator of the consequences that some theory assigns to various antecedent conditions" (Dennett, 1979, p.192) Ideally a simulation should accurately represent the hypothesis so that any conclusion based on the simulation is correct of the hypothesis. But the formalization imposed for formal tractability and other conditions in the simulation might contain factors not part of the "positive analogy" between target and model. Modelling_as_source_analogy vs modelling_as_thechnology ------------------------------------------------------- A source is a preexisting system used in devising a hypothesis. Mathematics are a good source. "Mathematical knowledge provides a pre-existing set of components and operations we can put in correspondence to the hypothesised components and operations of our target". Techonology is used to implement the hypothesis in a simulation. GOFAI: computer is a source analogy. BEER: computer is a technology. BIOROBOTIC MODELS: ------------------ Various ways in whitch robots and animals can be related as modelling. 1) Robots as source (Descartes, Weiner, infomration processing approach, etc.) 2) Animals as a source for robot construction. 3) Robots as simulations of animals or "how robots can be used to as physical models of animals to address specific biological questions" (Beer et al. 1998, p.777) Webb will focus on 3) with the following constraints: a) the system must be robotic (no computer-simulations) and situated in real world b) the robot must address a biological hypothesis. DIMENSIONS FOR DESCRIBING MODELS -------------------------------- The best model is the target itself, the identity model. From this point there can be deviations in 7 diferent dimensions. 1. BIOLOGICAL RELEVANCE: Is the biological target system clearly identified? Does the model generate hypotheses for biology? In this dimension models differ in the extent to with they intent to address and represent real biological issues. Another expression of this dimension is to distinguish between models as mathematical statements and their abstract analysis with no relevance to biology and models as claiming something about the real world. Mathematical-technological object <---> representation of biological system "The main criteria for relevance could be taken to be the ability of the model to generate testable hypotheses about the biological system it is drawn from. (...) The point is to use the robot model to make a serious attempt at addressing biological questions, at whatever level these may exist" (pp.1040-1) Relevant does not mean necesarilly detailed. Building a device for learning about reality is not the same as building an exact copy of reality. 2. LEVEL: What are the base units of the model? The level refers to which level the model pretends to represent: from molecular and membrane properties to social an population processes. The correct level might be problem specific. "For any organization, at any given level, its mechanism lies at the level below and its purpose at the level above" (Feibleman, J.K. (1954) Theory of integrative levels. British Journal of the Philosophy of Science 5:59-66, p.61) 3. GENERALITY: How many systems does the model target? The more target systems, the more general the model. A model must be relevant to be general but to be relevant it doesn't need to be general. What usually happens is that a model labelled "general" requires a large number of extra situation or task specific assumptions to actually get data from the model to compare to the observed target. Abstract != general. Membrane potential properties are very general but not too abstract, while logic reasoning is very abstract but not very general in biological systems. Generality must be tested for the targets, and depends on the nature of the targets. What is general is an empirical question. Webb suggests that it is better to work on concrete well-chosen specific models and then generalize from them. 4. ABSTRACTION: How many elements and processes from the target are included in the model? Abstraction is not directly related to level of modelling. "The amount of abstraction depends on how the modeller chooses to describe and represent the processes, not what kind of processes they represent." (p.1043) 5. STRUCTURAL ACCURACY: Is the model a true representation of the target? How well the mechanism in the model reflect the real mechanisms in the target, distinct from how the i-o behaviour of the system matches the target. "Accuracy of a model means there is "acceptable justification for scientific content of the model" (Rykiel 1996, p. 234) relative to the contemporary scientific context in which it is built; and that it is rational to attempt "experimental verficaction of internal mechanisms" (Reeke & Sporns 1993, p. 599) suggested by the model. Accuracy is diferent to relevance since it can be not trully acurate but still address (under certain assumptions) relevant biological questions. Nor accuracy is synonimous of "more detail". 6. MATCH: To what extent does the model behave like the target? "This dimension describes how the model's performance is assessed. It concerns testability: can we potnetially falsify the model by comparing its behaviour to the target?" (p.1044) There is difficulty about how match should be asessed and what the consecuences of a good match could be. Problem of UNDERDETERMINATION: The performace of similar behaviour is never sufficient to prove the similarity of mechanisms. 7. MEDIUM: What is the simulation built from? THE POSITION OF BIOROBOTICS --------------------------- The seven dimensions outlined above will be used to situate biorobotics in the space of biological models. 1. Relevance to biology. - Recent biorobotics is highly tied to biological research, there is inspiration, participation and feedback from biology. 2. Levels - Biorobotics integrates multiple levels in a bottom-up top-down interaction, because of "its emphasis on requiring a complete, behaving system as the outcome of the model". (p.1046) - The context of behaviour is always included. Biorobotics emphasise the physical level in the performance of sensing and action: the dynamics of interaction. 3. Generality - Without regrounding the generalisations by demonstrating the applicability of the results to some specific real examples, the problem modelled may end up being "biological" only in the terminology used to describe it." (p.1047) - Generality is sometimes better achieved from concrete examples to generalization than analyzing general problems which don't allow for the explotation of particular features. "Generalising a sensorimotor problem can result in changing the nature of the problem to be solved" (p. 1047) 4. Abstraction. - The abstraction in biorobotics is different than the abstractions in simulations. 5. Accuracy - Complete Robotic models are prefered to completely accurate ones, there is more lost on partial accurate models than complete interactive (but no so accurate) ones - Different robots with different levels of accuracy in different aspects give a good account if robust inter-robotic results are observed. 6. Match - It is currently difficult to asses biorobotic models. The level of accuracy makes difficult to match and compare specific data. - Biorobotic modelling guides biological research by driving the research of necesary mechanisms and other "necesary data" from biological research. 7. Medium - The most distinctive feature of biorobotics is the use of hardware. - Allows capturing real constraints and sometimes facilitates modeling by using identities in parts of the model: real antennae, air-borne plumes, real water (robotunna), etc. - Other advantages are: exploit true parallelism, physical preprocessing done by sensors. - The study of behaviour is also the study of the environmental circumstances and constraints. Simulating them implies that we already know which are the relevant ones. CONCLUSION ---------- Rather than biology as a source of ideas to build robots or robotic ideas as a source to explain mechanisms in biology. Robots as models OF animals: robots as a simulation technology to test hypothesis about biology. That is the aim of biorobotics. In the seven dimensional space "The strategy of biorobots has here been characterised as: - increasing relevance and commitment to really testing biological hypothesis; - combining levels; - studying specific systems that might illustrate general factors; - abstracting by simplification rather than idealisation; - aspiring to accuracy but concerned with building complete systems; - looking for a closer behavioural match; and - using real physical interaction as part of the medium." (p.1049) Different views of "models" reflect different views of the "the nature of explanation". ======================================================================== REMARCS ON THE DISCUSSION OPEN PEER COMENTARY ======================================================================== WHAT IS A MODEL --------------- Very interesting quote from the answer of Webb to Scheutz (who made a critique on lack of formal-logical description of mathematical isomorphic relationships to characterize a model as such.) << But if we cannot demostrate a partial isomorphism, what, asks Scheutz, "warrants the claim that a robot system is a model of an animal with respect to X"? I had no intetion of being covert in what I meant by "a robot system is a model of an animal". I mean the robot \emph{is being used} as a model of an animal. That is, I \emph{intentionally} conflated the "modelling relationship" with "modelling" \emph{qua} practice. To be explicit, I consider there be no "independent ontological" question as to whether "such a metaphorical relationship" actually exists between the two things. [1] That is, going beyonD Giere, not only is it the case that you "cannot eliminate the purposes of scientists form the evaluation of any model", it is my belief that you cannot eliminate their purposes from the question of wheter the system is a model at all. >> The discussion follows in a footnote marked as [1] in the quote above: << [1] This is not to deny that the choice of a metaphor or model, and its subsequent productivity or value, will depend in some way on the richness of the structural mappings we can make between their domains and their targets. The poin is we can construct and use untidy an cimplete mappings without ever being close to proving the existance of an isomorphism (partial or otherwise). There has been much interesting work recently on the theory of metaphor (Lakoff 1993) and the nature of analogical reasoning (Gentner 1983) that seems more applicable than logical model theory in understanding how we use models in scientific practice. >> ----------------------------------------------------------------------- INTERESTING PEER COMENTARIES ----------------------------------------------------------------------- * Effken, J.A. and Shaw, R.E. "An intentional dynamics approach to comparing robots with their biological targets" (pp. 1058) Supports Webbs paper and compared biorobotics with ecological psychology and provides methods to asses robot's performance based on affordance-effectivity matching. * Franklin, S. "Models as implementations of a theory, rather than simulations: Dancing to a different drummer" (p.1059) Abstract: "Robots, as well as software agents, ca be of use in biology as implementations of a theory rather than as simulations of specific real world target systems. Such implementations generate hypotheses rather than representing them. Their behavior is no predicted, but rather observed, and is not expected to duplicate that of a target system. Scientific knowledge is gained through the testing of generated hypotheses." Explains very interesting robot IDA that implements lots of neurological theories. * Giere, R.N. "The nature and function of models" (p.1060) "There are many different models representing different features to different degrees of accuracy. Which of these many possible models one wishes to consider depends on the purposes for which the model is being constructed. Thus, one cannot eliminate the purposes of scientistss from the evaluation of any model." (p. 1060) * Killen, P. R. "Doing versus knowing" (pp.1063-1064) Funny quote: " Requiring biological plausibility may tell us somehting about potential machinery at the next level, and it can improve our sense of what is plausible. But plausibility becomes less plausible the deeper it is pushed. The elephants supporting the universe may resto on other elephants, but as we descend it quickly becomes academic whether the next substrate is elephant or turtle. In many cases the cosntraint of plausibility merely serves, like the chalk marks on a tennis court, to keep the game interesting." (p.1063) "What robots are specially good for is sharpening Ockham's razor" (p.1064) * MacIver, M.A. "How building physical models can reduce and guide the abstraction of nature" (pp.1066-1067) MacIver explains how formal and computational abstraction required to make the model computable suffers a informational reduction where many essential constraints might be lost. Biorobotics, on the contrary, avoid this problem. * Metta, G. and Sandini, G. "Embodiment and complex systems" (pp.1068-1069) Action requires to be exploited to explain embodied and situated development and intelligence, active-perception been the fundamental developmental constraint it requires activer models (robots) to be adequately studied. * Steels, L. "The methodology of the artificial" (pp. 1077-1078) Explains how the methodology of the artificial is not only reduced to modelling (realistic) relationships but comparison of abstract artificial and real systems. "Romal models do not necesssarily describe a natural system. But they examine the impolications of certain assumptions that can then be used to understand natural systems. (...) unrealistic assumptions make it possible to investigate boundary conditions, isolate factors, highlight implications which would otherwise go unnoticed, perform demostrations by reduction ad absurdum, and so on." (p. 1078) The main points of the methodology of the artificial for Steels are: " 1) Artificial systems are developed in the first place to examine the consequences of certain assumptions, just as formal models are; 2) they require much more realism than formal models and hece privide much deeper insight; but 3) the goal is not to build realistic replicdas of natural systems. Their value for understanding nature lies in providing points of comparison with natural systems" * Young, D.L. and Poon C-S. "Soul searching and heart throbbing for biological modeling" (p. 1080) Provide a good discussion about top-down bottom-up interaction in complex system modeling. "In most instances, both top-down and bottom-up approaches may be needed in order to solve acomplex problem, and a model is fully validated only when bottom-up meets top-down". (p.1080)