Bad equilibria (and what to do about them)

I begin by arguing that the notion of economic equilibrium is an important analytical tool with which to understand the behaviour of today's networked computer systems. This is because the behaviours that such systems exhibit are in part a function of the preferences and desires of system participants; this gives such systems the flavour of an economic system. In economics, an equilibrium is a steady-state situation, which obtains because no participant has any rational incentive to deviate from it. Equilibrium concepts are arguably the most important and widely used analytical weapons in the game theory arsenal. The concept of Nash equilibrium in particular has found a huge range of applications, in areas as diverse and seemingly unrelated as evolutionary biology and moral philosophy. However, there remain fundamental problems associated with Nash equilibria and their application, which must be considered if we want to apply them to the analysis of computer systems. First, there may be multiple Nash equilibria, in which case, how should we choose between them? Second, some equilibria may be undesirable, in which case, how can we avoid them? In this essay, I will introduce work that we have done addressing these problems from a computational/AI perspective. Assuming no prior knowledge of game theory or economic solution concepts, I will discuss various ways in which we can try to engineer a scenario so that desirable equilibria result, or else engineer out undesirable equilibria.