PREVIEW.

Science.

... so the neural net selects one of several different strategies, based on how market conditions are observed and interpreted and predicted and classified, then further decides in that context how to trade. Different games need different strategies; therefore the overall strategy, the complete set of possibly conditional behavior is a mixed one. But not because there is another player with conflicting interests and strategies, who always has one strategy that can dominate any one strategy played by the net, whom the net therefore wishes to be unable to predict which strategy will be used next, not that. A mixed strategy is especially required in a quickly evolving market.

Q: So what you are proposing is making a system whose trading behavior consists of several distinguishable strategies? For sufficiently different market conditions?

A: Basically. Different indeterministic conditions have different factors interacting differently outside of player control and produce different random walks.

So imagine, then, as in game theory, the player doesn't know which one of several different games they are playing. Each game has as if different rules.

There are also gradients of rules. But not everywhere.

Smooth transitions between slightly different games, like moving from sheet to sheet in a stack of sheets of paper. But these transitions not everywhere.

If the system can classify situations into "gradient of classes present" or not, it can do better than simply assume each game occurs with some frequency and maximize over uncertainty.

That is something a neural net can do, if the logic underlying the method of classification somehow involves the weight coefficients arrived at in the network in deciding the bounds of classes.