About customized task about deep-symbolic-optimization HOT 1 CLOSED

Hey there,

I believe you want a custom task. In DSO, a Task is largely defined by a search space and a reward function. Both can be arbitrarily/custom designed. For symbolic regression (regression task), the search space is the space of mathematical expressions, and the reward function is fitness of that expression to a dataset. For the control task, the search space is the same, but the reward function is how well that expression performs when instantiated as a policy in an RL environment.

It sounds like you want to define another Task with that same search space (expressions) but a different reward function.

Can't say much else without knowing more, but take a look at the custom example we have and how to configure it. Basically, you'll want to (1) mimic part of the regression/control task (namely, inherit from dso.task.HierarchicalTask and have a dso.library.Library consisting of mathematical operators... the combination of these two essentially gives you the expression search space), then (2) implement reward_function() however you want.

Thanks,
Brenden

from deep-symbolic-optimization.