Theories of learning in economics
How should we model learning behaviour in economic agents? This thesis addresses this question in two distinct ways. In the first set of chapters the assumption is that agents learn through the observation of others. They use Bayesian updating which together with specific informational assumptions can generate the problem known as herding with the potential for significant welfare losses. In the final set of chapters the agent is instead modelled as learning by example. Here the agent cannot learn by observing others, but has a pool of experience to fall back on. This allows us to examine how an economic agent will perform if he sees a particular economic situation (or game) for the first time, but has experience of playing related games. The tool used to capture the notion of learning through example is a neural network. Throughout the thesis the central theme is that economic agents will naturally use as much information as they can to help them make decisions. In many cases this should mean they take into consideration others' actions or their own experiences in similar but non-identical situations. Learning throughout the thesis will be rational or bounded-rational in the sense that either the best possible way to learn will be utilized (so players achieve full rational play, for example, through Bayesian updating), or a suitable local error-minimizing algorithm will be developed (for example, a rule of thumb which optimizes play in a subclass of games, but not in the overall set of possible games). Several themes permeate the whole thesis, including the scope for firms or planners to manipulate the information that is used by agents for their own ends, the role of rules of thumb, and the realism of current theories of learning in economics.