The biggest problems I'm personally facing at the moment are midgame macro with imperfect information -- how many Gateways do I need to be producing from before I can safely take a third base or transition to endgame tech? -- and engaging Terran mech armies. It's very hard to know when or how to engage a Terran mech army, and you usually only get one chance to get it right. Dithering results in bleeding units you can't afford to lose.
The biggest problem in general is that rules-based approaches are asymptoticly approaching the limit of their capabilities. Either of the problems I've described are big undertakings but only scratch the surface of what PurpleWave needs to advance. Machine learning is the way forward. But so far real time strategy games have proved remarkably impenetrable.
Consider that professional players often place their initial buildings to provide Scarab-proof escape routes for their workers against Reaver drops that arrive ten minutes later. Such placements are only effective if you know to react to Reaver drops by funneling your workers through them to glitch the Scarabs. How are you going to let your bot learn that on its own? There's a lot of work to do.