multioptimizer

Multiobjective search over arbitrary data structures, with a monadic interface and multiple search strategies including Monte Carlo Tree Search.

TODO

• Distributed workers.

• Instead of specifying our objectives when we run the generator, it might be interesting to interleave maximize commands inside the do notation. For example:

  foo = do
    x <- choose xs
    maximize "obj1" (f x)
    y <- choose ys
    maximize "obj2" (g x y)

  The advantage here would be that we don't need to think about y to optimize obj1, which could hasten convergence. This seems powerful, but has a number of open questions I haven't thought through. What if we reach a different set of maximize commands on each run? What if we place a maximize command inside a loop, giving us n extra objectives? These seem like problems that can be solved with straight-forward extensions of the Pareto dominance idea, plus added bookkeeping, but I'm not sure if there are any applications that could benefit from this added complexity.