Things to do or try:

Don't bother trying every next. Pick the best half, and some random
subset of the rest. Use the time instead to explore more futures.

At search time, weight changes in objective functions by the magnitude
of the change, not just the number that went up or down.

Search by an iterated random process, where we look at all the first
steps, some of the second, fewer of the third, etc.

Integrate over the values of intermediate steps when evaluating a plan.

If we are stuck issuing a very low value move, jump back in time and
randomize.

Instead of motifs of fixed length, try sequences of random varying length.

Consider making a simple markov model of inputs rather than fixed motifs.

Allow lex ordering to include -byte in addition to +byte, for things
that decrease.

Allow lex orderings that treat their bytes as signed?

We should consider explicitly taking into account the tree of inputs,
since caching makes some kinds of exploration much cheaper than others.


when futures are bad in general, shorten them and have more of them.
When they are good, lengthen them and have fewer.


can use number of successful (better) backtracks as a way of gauging how
globally good my state is.


yolt = you only live thrice
hockey highlights: beginning of 3rd period!
     - after second goal, where's dude? ?!?! 63000 ish

prowrestling:  draw @21786
               then dq from being out of the ring too long
               then again.

wall street: "Have a great april fool's day!"

games:
%  tetris!
%  karate kid!
  contra!

  battletoads
  deadly towers


 