The Metagame approach to computer game praying, introduced by Pell, involves writing programs that can play many games from some large class, rather than programs specialised to play just a single game such as chess. Metagame programs take the rules of a randomly generated game as input, then do some analysis of that game, and then play the game against an opponent. Success in Metagame competitions is evidence of a more general kind of ability than that possessed by (for example) a chess program or a draughts program. In this paper, we take up one of Pell's challenges by building a Metagame player that can learn. The learning techniques used are a refinement of the regression methods of Christensen and Korf, and they are applied to unsupervised learning, from self-play, of the weights of the components (or advisors) of the evaluation function. The method used leads to significant improvement in playing strength for many (but not all) games in the class. We also shed light on some curious behaviour of some advisor weights. In order to conduct this research, a new and more efficient Metagame player was written.