I am writing a PhD proposal on a solution to an AI problem that hasn't been solved before (although there's related research). I am having trouble determining what kind of research questions / hypotheses are appropriate for this topic. I can't propose my solution is better, faster, more efficient than some previous technique (as is common in CS theses) as there is nothing to compare it to, it's a novel problem. I am not testing a hypothesis as to how something works (such as a theory of evolution, or quantum mechanics), nor developing some set of equations to describe some phenomenon. Rather I am using some standard machine learning algorithms to solve the problem. What form then should my hypotheses take? That I solve the problem to within some degree of accuracy?
1 Answer
Formal hypotheses are over-rated, but still often necessary for scientific communication, especially for a pedagogical exercise like a thesis proposal. The key question here is this: how will you know if you have succeeded in your intended work?
If the problem hasn't been solved before, then your core hypothesis is that it can be solved effectively with your proposed approach. From that spills out a whole bunch of secondary questions:
- Is the problem well-defined enough that "solved" is meaningful? This is often especially problematic in AI (e.g., "human-like behavior in a chat-bot", "good book recommendations")
- Relatedly, how can you measure whether you have actually solved the problem?
- What is the difference between a trivial/toy solution and an effective solution? Again, a big problem in AI, where it's often possible to make systems that technically solve the problem but practically do not (e.g., for real applications they require unavailable curated data, or a thousand years of training time).
- What sort of "partial solution" milestones are there along the way that will help you tell whether you are on the right track?
- What would an optimal solution be? Can it even be identified?
- What will you learn if your chosen approach doesn't seem to be working? Will it be possible to definitively tell that it cannot be a solution?
- How will you know when you have done enough?
If you can provide good answers to all of these questions (or even just most of them), then you will have a very solid thesis proposal, built around the hypothesis of "Approach X can solve Problem Y."