Knowledge-Free and Learning-Based Methods in Intelligent Game Playing by Jacek Mandziuk (Springer, 2010) is not a book I intend to read from cover to cover. It explores computational intelligence-based methods (as opposed to AI methods) for playing such games as chess, checkers, backgammon, and Go. What caught my attention was a section entitled “Training with External Opponents.” In this section the author summarizes Susan Epstein’s paper “Toward an Ideal Trainer.”
The most important conclusion of her research is that training with one opponent, whether in self-play or against an expert player, is insufficient. If a player is trained by a single expert, that player will not necessarily be competent when confronting weaker or unconventional opponents. The trainee needs a diversity of trainers. “In order to develop a flexible and ‘intelligent’ player (i.e. the one capable of passing the Turing test in games) it is necessary to confront it with a large range of playing styles and various playing skills.”
She initially tried to improve upon single-opponent training by adding some noise into the decision-making process of the trainer. This strategy brought better results, but “it still does not guarantee that the training is sufficiently universal and that the part of a state space visited by the learner is representative enough to allow efficient play against intelligent, nonconventional opposition.” (p. 126)
Is it valid to extrapolate from the world of CI training when assessing the value of expert trading mentors? In one obvious respect the extrapolation breaks down: computers don’t have biases, they don’t carry psychological baggage, and presumably they all have the same willingness to learn. But if we overlook that “minor” problem, I think there is merit in the claim that a trainee needs a diversity of trainers. Assume that the mentor/trainer is truly skilled at a certain style of trading, that he is also a skilled teacher, and that his style of trading fits the personality of the trainee. Even then, at the end of the day (and I don’t mean that literally) I doubt that a flexible, intelligent trader emerges. The trainee has learned one style of trading; that makes him neither flexible nor intelligent. Indeed, if compared to a sophisticated algorithmic computer program he might seem decidedly unintelligent. When the market throws him a curve ball will he be prepared to deal with it? If the tone of the market starts to shift, will he be flexible enough to go with the flow?
I am not suggesting that the would-be trader spend many thousands of dollars to study under several mentors. With some very good blogs, webinars, and (as always) books available, the trainee can become familiar with a range of styles. Admittedly, this kind of learning suffers from being passive. A web surfer doesn’t have the advantage of a teacher who prods, corrects, and praises. If he needs this interaction he has a range of alternatives, from self-talk to structured chat rooms to full-blown educational programs.
It goes without saying that training needs to be complemented with practice, not only for would-be traders but for would-be game-playing computers as well. Epstein proposed “lesson and practice” learning that consists in “interleaving the periods of training with strong opponents (the lessons) with periods of knowledge consolidation and usage (the practice).” (p. 126)
In the meantime, an interesting exercise for the self-taught trader might be to imagine three or four trainers with very different styles together in the same room, trading the same underlying instrument. Would A be taking the other side of B’s trade? Would C be waiting for two or three setups a day while A was clicking away? Would B scale out while C held a full position to the end? Who would be squeezed and when? What would these trainers be saying to one another, expletives not necessarily deleted, during the course of the day? If you overlaid (and color coded) their trades on a single chart, would you know a little more about market dynamics than you do now?