This reminds me of a topic that came up in Thoery a week or so ago.
Let's ignore the difficulty of creating a game-theoric type perfect strategy, and just assume you've got one. The question is when you use it vs. other exploitive sub-optimal strategies. I believe that an elegant solution to that very problem, on numerous levels, is represented in the Iocaine Powder Roshambo bot.
http://ofb.net/~egnor/iocaine.html
It's got both a game-theoretic strategy (random play) and exploitive ones (history matching & frequency analysis). It then has two levels of meta-strategy to select between them. I'll let you read the details. I believe a similar meta-strategy could be adopted in poker. I don't know, however, the degree to which a player could mentally do the score keeping necessary to actually implement it.
Might be more of a feel thing to really make it happen. But at least knowing what you're trying to approximate could be valuable.