Designing and Training a NeuralNet for trading algos

i've recently gotten the machine learning / neural net bug and have decided to take the plunge.

I have a set of trading strategies (mostly uncorrelated), mostly technical analysis based, where the base data is footprint data (price - qty bought - qty sold) at short intervals...about 5 seconds (so, not ultra HFT, but still pretty active in the day-trading space). Some of these strategies are close to linear (linear of log(x)), others are not. I'm in the process of building a neural net and training architecture, in an attempt to steer the ai to make more robust (and automated) versions of what i'm already doing manually.

As anybody who has played with neural nets knows, one of the problems is when the system finds a local minimum, it stops looking. So, if the type of strategy that i want the NN to research is not in the 1st local minimum found, then the NN never gets a chance to study my strategy.

Does anybody here have any automated strategies for teaching the NN, "how to teach the NN to learn...learning how to learn" (neural net hyperparameter optimization). Automating the search for the best hyper params (how many layers should i use...how many nodes per payer...step interval...how to best pick random staring sequences, etc..). I'm aware that this is the current research space for ai (with lots of data, the computing resources grow exponentially, and so its a bit of brute force and random).

Anybody have any brilliant ideas i can borrow/steal?