Choices Between Simple and Compound Lotteries: Experimental Evidence and Neural Network Modelling

An experiment on choices between single and compound lotteries is presented, and results are calibrated with neural network models. Many subjects tend to average out probabilities, though behaviour becomes more rational with more exposure to compound lotteries in the practice stage. The Prior Knowledge Model hypothesizes that subjects categorize stimuli according to the prior knowledge acquired in their long-run learning history; practice stage cues help them referring to the relevant learning history. The trained networks predict the behaviour of about 3/4 of the subjects with transitive preferences; the model can explain where we would expect the trained networks to fail.