Modelling Science

Institutionalized science is a social process in which theory production and evaluation is distributed between agents. It can be seen as implementing a search algorithm trying to maximize empirical adequacy in some theoretical space. Following David Hull’s Science as a Process (1988), this algorithm
is a ‘genetic’ one, where objects such as ideas, concepts, theories, etc., are produced, transmitted and selected in a darwinian fashion. We aim to model Hull’s theory of the scientific process as a multiagent system and work out its consequences using computer simulation, allowing exploration of the conditions under which the institutionalized science is most efficient.

According to Hull, Darwinian evolution is a general process. It requires interactors and replicators. As long as the actions of the interactors cause the differential reproduction of the replicators, you have evolution. Organisms and genes implement an instance of that process, but so do scientists and ideas. To propagate their ideas, scientists must acquire credit, i.e. the consideration of their peers, with whom they are both competing and collaborating. Citation of someone’s work in a paper is a mechanism that trades credit for support. Sharing credit with your students provides something akin to “inclusive fitness”, since they are vectors for your own ideas.

In a multiagent model, agents are usually heterogeneous. They can be more or less creative, communicative, sociable, aggressive, generous, etc. These characteristics are expressed as real numbers and taken into account by the formal rules of behaviour implemented by the agents. Agents are also situated in an environment where interactions are mostly local. In our model, this environment is a social network: scientists interact with their students and their collaborators. They also write peer-reviewed articles that allow them to share their ideas with the whole community. These ideas are strings of information that can be formalized as vectors in a multidimensional space. When ideas are transmitted from one scientist to another, some bits of information can be mutated, introducing noise in the transmission. Each idea is assigned some “empirical adequacy” value, using an arbitrary objective function. This function constitutes the problem space that the agents are
working in. They do not have direct access to its values, but they can try to approximate them using “tests”. Overall, we can measure the efficiency of the system by how close the subjective ratings given by the agents to their own ideas are to their “real”, objective values.

PayetteThe complexities of this model would make it analytically intractable, but simulation allow us to examine, one step after another, what happens in the system, and to see the effect that different parameters have on its efficiency at maximizing empirical adequacy. It is conceivable, for example, that too much interaction between scientists would lead the system to settle at a local optimum, while having small, close-knitted but independent communities would allow a more thorough exploration of the problem space. Different results here would suggest different norms as to how science should be socially organised.