Consider a situation such as that of Exercise 1.1, i.e. the first exercise at the end of chapter one. We imagined that we were going to get lots of pictures of half naked women and about ten times as many trees. Your job was to classify a new image. Now you already have odds of 9 to 1 that it's a tree. Shouldn't you use this information in doubtful cases? In this situation, you are treating the models themselves as objects on which to put probabilities. Doing this systematically gives the Bayesian solution, which is an attempt to rationally utilise other sources of information about any one model being the `best'.
More generally, we may have prejudices about the world based on other experiences upon which we are inclined to rely more heavily than a single small amount of data. In the case of the coins, I may have a predisposition to believe that the probability of Heads is much closer to 0.5 than the 0.8 which this single experiment would suggest. This may reflect my having thrown very similar looking coins more than ten times in the past and having got much less than 80% of them coming Heads. I do not approach this experiment in a total vacuum. These considerations lead to Bayesian ideas about how to choose the `best' model.