The most important aspect of the proposed approach is that it helps to elucidate the mechanisms underlying various mental properties. This is possible because many of the basic features of the control system reaCog are known. Using the words of Cruse & Schilling (this collection), it constitutes a “quantitatively defined system”. As all components are realised as artificial neural networks, all information about the number of neurons, the connection weights between them, and the way individual neurons process information is available. More importantly, however, the basic functional architecture of the system is well understood. Which modules are connected in which ways to other modules, how they receive their input, and what other parts of the system might be affected by their outputs does not have to be figured out by painstaking investigation—as is the case in biological research. Because these facts about reaCog are known, it is possible to provide detailed mechanism descriptions. In this way, reaCog’s ability to plan its future actions by internal simulation can be explained by reference to the interaction of its various sub-modules: a problem detector is activated when sensory input indicates that current behaviour will lead, if continued, to adverse effects for the system (e.g., falling over). This leads to the abortion of current behaviour and activation in the Spreading Activation Layer, which randomly excites the Winner-Takes-All network (WTA-net). After some time, the WTA-net adopts a relaxed state in which only one of its units is active. This active unit in turn stimulates its counterpart in the Motivation Unit Network, leading to activity of the corresponding reactive procedures. These provide motor output that can be redirected to the body model, which then simulates the execution of the proposed behaviour and predicts its likely consequences. If the system predicts that the problem will persist, the process of internal simulation goes on until a solution is found, which can then be used to control the actual movements of the system.
Explanations like these contain a lot of information about which functional subparts of a system are engaged during the exercise of the ability in question. In this particular case it makes clear how the ability to plan ahead, a cognitive ability, depends heavily on basic reactive structures that are designed to control specific leg movements as well as an internal model of the body. The same is true for various other capacities like attention and Theory of Mind. Thus, new insights into the mechanisms responsible for those phenomena in humans could be gained by considering how body models and motor control mechanisms are realised in our case and how these systems interact. In other words, the bottom-up approach may lead to new directions for future research concerning human psychology by suggesting how specific functional modules interact in order to bring about a particular target phenomenon. Whether this approach is tenable depends on the degree to which findings pertaining to the artificial system might legitimately be used to draw conclusions about human beings. I will propose a number of constraints to ensure that this condition is fulfilled below.
Another class of questions that a bottom-up strategy is well designed to answer has to do with the evolution of cognitive capacities: how did cognitive systems evolve from purely reactive systems? How did emotions, attention, or even consciousness arise? What are the natural precursors of these phenomena? Cruse & Schilling (this collection) show convincingly that no completely new neural modules are needed in order for such properties to occur. Rather, minor changes in the basic architecture might suffice to generate radical extensions of the abilities of a system. In this way, a reactive system with a body model can acquire the ability to plan ahead if it is able to disconnect its motor system from the physical body and instead send the motor signals to its internal body model. No novel “planning module” is needed. Already existing modules just have to become dissociable and can thus acquire new functions (Cruse 2003). In addition, the target paper suggests an answer to the question of the evolutionary function of cognition understood as the ability to plan ahead: it was the necessity of being able to control a complex body in a complex environment that made this ability highly valuable. Detecting problems by perception, finding innovative solutions by internal simulation and acting on them are capacities that are extremely advantageous for any organism possessing a body with a high number of redundant degrees of freedom (see Cruse 2003). This is in line with, and actually extends, the widespread assumption that the evolutionary function of cognition is to deal with environmental complexity (Godfrey-Smith 2002).
In the previous paragraph we saw that the framework Cruse & Schilling (this collection) present is well-equipped to give new insights into the underlying mechanisms of psychological phenomena and the evolution of cognition, as well as a promising approach to creating highly flexible and intelligent robots. There are, however, some problems the proposed strategy has to face, especially if the control structures become increasingly complex. I therefore want to suggest a set of three constraints on good bottom-up explanations of biological/psychological phenomena.
Adequate matching criteria:[3] At two points the research strategy described in section 1 involves a comparison between the behaviour of an artificial system on the one hand and a biological system on the other. First, this is the case in the development of neural network models of animal behaviour. In this context, the comparison is used to ascertain whether the proposed model of the mechanisms underlying certain capacities (e.g., walking) really reproduces the original behaviour of the animal (e.g., a stick insect). Second, there is a similar process of comparison involved in the application of psychological concepts to the complete system. At different points in their discussion, Cruse & Schilling (this collection) argue that their system has certain mental capacities because it exhibits behaviour (or would exhibit it if certain extensions were implemented) connected to those mental capacities in humans. So, for example, just as the performance of athletes might worsen if they consciously attend to what they are doing, the activation of the attention controller in reaCog can lead to poorer results compared to unimpeded execution of the reactive procedures.
Both processes of comparison require criteria to identify when the behaviour of the artificial system and that of the biological system are relevantly similar, i.e., similar enough in order to provide evidence for the claim that similar mechanisms are at work in both cases or that the artificial system and the biological system share certain psychological characteristics (Datteri & Tamburrini 2007). The difficulty of finding such criteria increases the more the bodies of the compared systems differ. In some cases they might nonetheless be easy to find and relatively uncontroversial. This, however, is not always the case. For instance, in their discussion of emotions—and more specifically the emotion of happiness—, Cruse & Schilling (this collection) suggest that by increasing the threshold of the problem detector reaCog would take more risks, thus behaving similarly to humans when they are happy. Now, the question is whether the kind of risky behaviour exhibited by reaCog when the threshold of its problem detector is increased is the same kind of risky behaviour humans exhibit when they are happy. Only if this condition is fulfilled can the similarity be taken as evidence that reaCog shows aspects of the emotion of happiness.
Biological plausibility: Any proposed mechanism should be biologically plausible, i.e., it has to be reasonable to assume that the capacities of the organism that we are trying to understand are really based on such a mechanism. This can, at least to some degree, be ensured by trying to create similarities between the artificial and the biological organism on a basic structural level, for example by using artificial neural networks. Furthermore, it is necessary to decide how fine-grained the model should be. Should the model take brain structures, neurons, or subcellular elements as its basic building blocks? Should intracellular processes be neglected or are they important? The answer will of course always be relative to our particular epistemic goals. Finally, there are different options regarding the way artificial neurons process information, i.e., how they calculate their output value depending on the weighted sum of their inputs. All these factors might turn out to be important if the results are to be used to infer biological mechanisms.
The requirement of biological plausibility shouldn’t, however, be overemphasised. Cruse & Schilling (this collection) stress that they are not trying to present a realistic model of neuronal activity in living organisms. Hence, they are using biologically implausible, non-spiking artificial neurons as the basic elements of their architecture, while noting that some authors (referring to Singer 1995) have located the neural basis of consciousness in synchronously oscillating spikes. This, however, is not a weighty objection to the proposed approach since it is designed as a functional approach. The question is: how do different functional subsystems like a system for controlling the swing-movement of a leg, a system modelling the robot’s body, and a system allowing for the selection of different internal states interact in order to produce certain emergent phenomena? Therefore, the concrete physical realisation of these subsystems is of only secondary importance.
Transparency:[4] Doubts about the strategy of using artificial systems in order to understand biological systems arise because even if we were to create an extremely intelligent robot, it would not necessarily help us to understand the mechanisms underlying its intelligence. Rather, we might be faced with yet another complex system whose workings we do not understand (Holland & Goodman 2003). Now, the approach Cruse & Schilling (this collection) present is specifically designed to discover emergent properties, i.e., properties that were not explicitly implemented. This means that there will be a high risk of finding properties in the complete system that cannot be readily provided with a clear-cut mechanistic explanation involving the cooperation of the system’s components. Although the explanations of the occurrence of various psychological properties presented in the present paper are quite convincing, the bottom-up strategy might eventually exhaust its potential if the complexity of the system is further increased.