Come up with more ideas for policies that are structurally nontrivial

[timql]

Leave ideas, links and etc. here!

[timql]

One idea: evaluation functions for games, or maybe even an entire move function (so including game tree search/minimax or something). It'd be kind of a backwards way to train a neural net to play a game since neural nets can already play games without relying on explicit evaluation functions (i think), but nevertheless I think it's a fairly good example of a nontrivial policy.

[timql]

More ideas:

• Auction pricing -- the function could both approximate a policy function itself and predict new prices (e.g. airline prices)
• A wrapper that just trains a neural net in the background each time the function is called to potentially identify patterns and outliers in the function's inputs (fraud detection).
• AI agent in some environment, like an RPG where its state can be represented numerically (health, number of items, etc).
• Bank pricing policies -- mortgage/interest rate/stock pricing (where would we find examples of these rules?)
• Travel times (approximating, for example, a Google API query)

[ch3njust1n]

Software project planning time estimation