Reviewed by
Bruce Edmonds
Manchester Metropolitan University
This book presents "Agent_Zero" a minimalistic model of agent decision making, that is then thoroughly explained and explored. This model includes social, emotional and rational elements and aims to show how this can generate a range of observed socially-situated behaviour. The books starts with a useful summary of the book’s thesis. Then Part I motivates, presents and explores the core of Agent_Zero. In doing this it covers a lot of philosophy, neuroscience and social science. It explains the model with admirable clarity and gives a flavour of its characteristics. Part II presents a series of simulations based around this model showing how a number of different simple scenarios can be obtained. This part of the book is very reminiscent of Growing Artificial Societies (Epstein and Axtell 1996) but with a relatively small number of agents and looking at each case in more detail. Part III considers several possible extensions, giving illustrative simulations using each to show their potential. Finally, Part IV briefly considers future research directions and concludes. Again, the book helpfully provides appendices giving the maths, Mathematical and NetLogo code as well as parameter settings for the various exhibited runs. All the code plus additional material (such as videos) is freely available on an associated website.
Due to its rich and varied content, assessing the quality of the contribution is not straight forward. In this regard, the book provides two, very different, pictures of itself. Firstly, as in the title, we can anticipate neurocognitive foundations for generative social science. However, later, the exhibited simulation runs are described as "closer to parables than to mature scientific claims of any sort [...] they are Computational parables – fables if you prefer" (p.90). The former gives an impression of solidity and progress, the latter something you might tell children to help them sleep.
Let us start with the foundations. There are lots of pictures of neuroanatomy and interesting discussion of what is, and might be, going on in the brain. This part of the book is excellent as a mini-introduction to some of the thought and knowledge in this area. However, on careful reading, the only part of all of this that ends up in Epstein’s model is one equation for conditioning, the Rescorla-Wagner (1972) model that derived from experiments on conditioning and not from studies of neurocognitive architecture. This disconnect between the discussion of brain structure and the content of Epstein’s equations is underlined by his comment that "if my account of the tissue science is wanting [...] I do not believe that per se endangers the mathematical model" (p.4).
The Rescorla-Wagner model encodes some simple observations about learning: that organisms learn when events violate their expectations; and that the change of expectation depends on the difference between expected and observed levels, so that learning (or forgetting) will be rapid to begin with followed by a period of asymptotically approaching the observed level. It has become a sort-of standard in experimental psychological fields. Its use as a “foundation” in this book suggests that it is something we can rely upon. However, as a retrospective review of the model and its influence (Miller, Barnet and Grahame 1995) summarises it:
"It is concluded that the model has had a positive influence on the study of simple associative learning by stimulating research and contributing to new model development. However, this benefit should neither lead to the model being regarded as inherently "correct" nor imply that its predictions can be profitably used to assess other models".
The most interesting aspect of Epstein’s core model is that it attempts to integrate social, emotional and rational components of decision-making. It does this in the following manner: (for any particular binary decision) each of these components is represented as a separate real number, these are added up and if the total is bigger than a threshold then the action is taken. Thus, this assumes that these three aspects are independently processed and then additively combined. These are staggeringly bold assumptions, which, unfortunately, are not supported by neurocognitive (or any other) evidence. Epstein summarises his contribution as "My general claim is that wherever emotional, deliberative and social components combine to generate behaviour, the agent_zero framework can apply" (p.17). However, the strong assumptions of independent processing and additive combination militate against such a general applicability[1].
So let us turn to the simulation end of this exercise. As Epstein says "I am less interested in the accuracy of the components than in the generative capacity of synthesis" (p.3). Of course, when Epstein talks about the exhibited simulation runs being like fables, he is not seeking to denigrate his own work, but is rather pointing to the fact that the simulations are abstract explorations – simulations that are not empirically validated[2]. What he presents is a series of scenarios concerning three agents interacting within a 2D field of locations (each of which has an 1D intensity). These, it is suggested, could be related to many things, including: Leadership, Health Interpretations, Fight vs. Flight, the deliberation of juries, the Arab Spring, Spirals of Mutual Escalation, and Revolution.
Epstein presents some defence for this lack of validation. He says: "Scientific theory should not aim at realism. [...] There are no ideal gases or frictionless planes. But these limiting cases [...] turn out to be the productive theoretical entities. [...]Agent_Zero is such an idealisation" (p.192). However, these established models approximately predict observed outcomes in many circumstances and can be corrected to make almost completely accurate predictions in many others. There is no evidence that Agent_Zero has similar powers. Rather, we just have to presume that Epstein’s model may be as successful as these. He also suggests his models might guide future data collection, saying "where data do not yet permit, theoretical models like Agent_Zero can guide its collection" (p. 185), however, again, this presumes that Agent_Zero is going anything like in the right direction.
Here, I think, Epstein makes a mistake that many do – it assumes that simpler models are somehow more general. The reverse is true – one can make a model less general by adding detail into it – but this way around is a fallacy. To see this, just think of eliminating all variables but a constant from an equation – this is simpler but unlikely to be correct or useful for anywhere but a few points where the constant happens to be correct. In other words, the simpler model can be much less general. The reason is common sense: when simplifying we may leave out aspects or processes that are essential to the representation of the observed phenomena. Whilst there is dispute about what the best strategy for arriving at useful models is[3], many simple models have turned out to be either wrong or deeply misleading[4]. Here, it appears, that Epstein is assuming such a simple model can be applied somehow, almost separate to any evidence. It is true that models can never be completely realistic, but the stronger the relationship with reality the better.
To sum up, this book proposes a bold synthesis between cognitive or neuroscience and social science. It is an interesting read, bringing together the thought from many influential thinkers and different fields. Epstein is surely right that neurosciences and social sciences need to be bridged (and that agent-based simulation might provide such a bridge). He does make some interesting suggestions (e.g. that some important social influence phenomena may happen via a contagion of expectations rather than of behaviour). The models presented here could be pointing in the right direction, and thus provide guidance for future theoretical and empirical research, but this book provides little reason for believing that this is so. It has no foundations other than a single equation for conditioning and no direct validation of its results. Thus, whilst the book provides some interesting ideas, it is unclear that it provides a good starting place for subsequent exploration.
Notes
2 With the exception that it fits the Latané and Darley (1968) experiment which showed that a person would take longer to vacate a room filling with smoke if there were others there who were ignoring the smoke.
3 The, so called, “KISS vs. KIDS” debate, which is far from settled.
4 Of course, many complicated models have been equally as unsuccessful – it is just that simplicity is not a guide to truth – either way (Edmonds 2007)!
ReferencesEPSTEIN, J.M. and Axtell, R.L. (1996) Growing Artificial Societies: Social Science from the Bottom Up. MIT Press: New York.
LATANE, B. and Darley, J. M. (1968) Group inhibition of bystander intervention in emergencies. Journal of personality and social psychology, 10(3), 215.
RESCORLA, R. A. and Wagner, A. R. (1972) A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. Classical conditioning II: Current research and theory, 2, 64-99.
MILLER, R. R., Barnet, R. C., and Grahame, N. J. (1995). Assessment of the Rescorla-Wagner model. Psychological Bulletin, 117(3), 363-386. http://dx.doi.org/10.1037/0033-2909.117.3.363
Return to Contents
of this issue