Order this book
Each chapter ends with simulation exercises that replicate the findings presented in single papers. The book provides web access as well as the files necessary to run the simulation exercises. The programme is user-friendly, and no programming skills are necessary to develop a connectionist simulation model with this software. The simulation exercises themselves supply clear and simple instructions for getting acquainted with the FIT program. Moreover, the book contains a manual providing essential information about the program. The software is not intended for multi-agent modelling, but rather to represent mental processes. In other words, it is envisaged for simulation in psychology. The combination of research articles and exercises makes the book an excellent basis for lectures at graduate-level seminars and courses, introducing students of psychology to the world of simulation.
The book is organised in four sections: each consists of between one and four representative papers. Part one offers an overview of the connectionist approach and recurrent and feed-forward models used throughout the book. The second part discusses how belief in causal strength is acquired when people observe multiple co-variations between causes and effects. Part three investigates the acquisition and generalization of (group) stereotypes and biases and person impressions. The final part four addresses the question of how attitudes are formed when we pay attention to persuasive messages.
In the first section, hebbian and delta learning algorithms are introduced. In particular, essential properties of the delta learning algorithm are described. These properties are utilised in models described in subsequent sections to generate the phenomena in question. First of all, the adjustment of connection weights (i.e. learning) can be illustrated by the most simple and straightforward case: cause A has effect E. A has thus an activation level of 1. However, by default the connection weight is set to 0. The network does not predict effect E. Dependent on the learning rate, connections between A and E are then gradually adjusted. This leads to a sample size effect: the greater the number of cases in which A and E occur together, the more robust the predictions of the network become.
The next case subjected to scrutiny is causal competition. For instance, A and B may always co-occur with effect E. However, while A occurs every time together with E, B occurs only, e.g. every second time. Hence A is a stronger explanation of E than B. The learning process of the algorithm leads to a competition of A and B for the available connection strength, which is limited to 1. Because of this competition, the strongest cause tends to block the acquisition of other causes.
Finally, a cause (input unit) C might influence several effects (output units) A, B, D, E etc., e.g. the symptoms for a disease or a friendly character. This leads to a weakening of the connection weights. Hence, we observe a diffusion of the causal influence (i.e. connection weights) of the input activation to multiple output nodes. In contrast, such a diffusion does not occur in the case of one single activation of the opposite trait (e.g. behaviour expressing aggression). Such lack of diffusion leads to better recall for inconsistent information.
These simple properties can generate a number of phenomena that are traditionally investigated within the remit of social psychology. Interestingly, all the models described in this book are validated by the means of classical statistics. This circumstance is seldom found in social simulation. In psychology, there is no gap between modelling and statistics.
This review is not the place to present a comprehensive report of the book's findings. Instead, selective examples can be described, taken from each section.
In the case of causal attribution (part 2), connectionist modelling is able to replicate the psychological finding of an attribution bias: people tend to ignore relevant causal factors, leading to an oversimplification of the causal perception. Connectionism explains this psychological finding by invoking the blocking effect in the causal competition. In cases of stereotype formation (part 3), a mechanism is described that can generate emergent attributes: Think of a Harvard-educated carpenter. Such a person might be assumed to be non-materialistic, an attribute not highly salient in either the Harvard or carpenter stereotype. However, while Harvard students are qualified for high-paid jobs, carpenters are typically low paid. This could lead to the inference that a Harvard-educated carpenter might have a non-materialistic attitude, simulated by a pattern of three stereotypes: ABC, DEF, and CFG. Given a stimulus pattern ABDE, a model that only applies a single knowledge structure could infer C (using stereotype 1) or F (using stereotype 2). More flexible models might infer C and F by using both stereotypes. However, the connectionist model is also able to infer G, present in neither pattern 1 or 2. This outcome was achieved by inferring C from stereotype 1 and F from stereotype 2. From this inferred characteristics C and F, the model then leads to the inference of G. Thus if, for instance, A is the stereotype of a Harvard student and D the stereotype of a carpenter, G might be the stereotype for non-materialism.
Other models in this section simulate discrimination: the well-known phenomena that a minority group is seen in more negative terms despite the fact that the proportion of positive and negative items is identical in minority and majority group. This is due to the sample size effect: while the larger sample of the majority group de-emphasises episodic (negative) weights, such information is recalled more easily in the case of a minority group.
In part 4, a model of heuristic reasoning is described: for instance, an attitude object (car) might be connected to three cognitive nodes (fast, dry, polluting). These cognitive nodes are then related to evaluative nodes (good, bad). During the learning phase of the network, evaluation nodes are activated. In this example, the positive evaluation (fast, dry) finally outweighs the negative (polluting) evaluation, leading to a general positive attitude towards car driving. Positive and negative evaluation provides an example of competition.
These are only a few examples of how social psychology draws on the properties of the delta learning algorithm. The book demonstrates how biases and shortcomings of human reasoning can be explained by the single mechanism of connectionist learning. This is a considerable achievement in terms of explanatory simplicity. On the other hand, it may be questionable whether this account is too over-arching: if everything is reduced to the same explanation, it no longer becomes possible to discriminate between individual phenomena. According to a Popperian view on science, an explanation for everything is an empty one. Thus, social-psychological connectionism risks to appear as something like a tautology.
Moreover, the delta learning algorithm is empirically confirmed in particular by findings from animal conditioning. It might be questioned whether this is sufficient to explain the mental capacities of human beings: What is the crucial difference between humans and other species?
These questions, however, do not undermine the value of the book: they simply illustrate how it stimulates further thinking.
Return to Contents of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2008