One thing you can try to do with very hard problems with very sparse rewards is to suppose a slope of incremental rewards which lead toward the goal. Then you d

One thing you can try to do with very hard problems with very sparse rewards is to suppose a slope of incremental rewards which lead toward the goal. Then you don't have to figure out how to get to the goal. You can just collect the rewards on the slope. Figuring out how to take the next step is much easier to do than figuring out the full path. So its a much more tractable problem. It does require treating problem-specific advice as if it were true, since that lets you infer a gradient you haven't figured out for yourself. Another approach that can be taken when dealing with hard problems is to start with the easiest possible version of that problem and solve that problem. Then to complicate it ever so slightly and solve that problem. Through doing this repeatedly you can move along a path of increasing understanding related to the problem you are solving. Upon reaching the most complicated problem which you struggled with previously, the problem is now an incremental addition of complexity to a problem you already know how to solve, rather than a problem which you don't know how to solve. This reduces the amount which needs to be learned to make a successful step. Which makes the problem of taking that step much easier.