[reference] — from The Craft of Research and standard scientific-method guidance. The hinge between “an area I’m curious about” and “an experiment I can actually run.” Finding a problem got you a candidate gap; this note sharpens it into something precise and testable.

The kid version first

“I want to know about drones” is a wish, not a question you can answer. You can’t tell when you’re done, and you don’t know what to measure. Research needs you to grind the fuzzy wish down to a sharp question you could actually check — like turning “are dogs smart?” into “can my dog learn to sit in fewer tries than my cat?”

Look what the sharp version gives you for free:

  • What to measure — number of tries to learn “sit.”
  • What to change — dog vs cat.
  • When you’re done — when you’ve counted the tries for both.
  • A clear answer — either the dog took fewer tries, or it didn’t.

The whole job of this note is that grind: fuzzy topic → sharp question → (often) a hypothesis you could prove wrong. Get this right and the experiment designs itself; get it wrong and no amount of clever work saves you, because you never defined what “answering it” means.

The funnel: topic → question → hypothesis

  • Topic — a broad area (“drone landing,” “LLM reasoning”). Too big to answer; it’s a direction, not a target.
  • Question — a specific, answerable thing about that topic (“does controller X reduce landing error vs standard PID in wind?”).
  • Hypothesis — a predicted answer you can test and that could turn out false (“controller X reduces landing error by ≥20% at wind >5 m/s”).

Not every project needs the third step (see exploratory research below), but every project needs the second. Beginners stall at the topic stage and call it a research plan.

Operationalizing — turning fuzzy words into measurements

The critical move, and the one people skip: every vague concept in your question must become something you can actually measure. A concept you can’t measure isn’t a research question yet.

Vague wordOperationalized as…
“better”+X% accuracy on dataset D / −Y ms p95 latency
”robust”performance degradation under noise level N
”efficient”energy per operation / FLOPs / cost per 1k requests
”learns faster”epochs (or samples) to reach threshold accuracy

This is called defining your metric, and it’s a choice with consequences — “better” measured by accuracy vs. by latency can give opposite answers. Pick the metric that actually reflects what you care about, name it explicitly, and justify it. A huge fraction of weak papers are weak because “better” was never pinned to a number.

Variables — name the four kinds

Any experiment is really “change one thing, measure another, hold the rest still.” The vocabulary:

  • Independent variable — what you deliberately change (the controller: X vs PID). The cause you’re probing.
  • Dependent variable — what you measure (landing error). The effect.
  • Controlled variables — what you hold constant so they can’t interfere (same drone, same payload, same test area).
  • Confounding variables — the sneaky ones: something that varies along with your independent variable and could be the real cause of the effect. (If you tested controller X only on calm days and PID only on windy days, “weather” is a confound — you can’t tell if X won or the weather did.) Hunting confounds is the heart of valid experiment design.

Hypotheses & falsifiability

A real hypothesis makes a risky, specific prediction that could turn out false. This is Popper’s falsifiability: a claim is scientific only if there’s some result that would disprove it. “Our method works well” is unfalsifiable fluff (what result would count as not working?). “Our method achieves ≥90% accuracy on benchmark B” is falsifiable — run it and find out.

The formal framing (used with statistics — analysis):

  • Null hypothesis (H₀) — the boring default: “no effect / no difference” (controller X is no better than PID).
  • Alternative hypothesis (H₁) — what you suspect: “there is an effect” (X is better).
  • You design the study to see whether the evidence lets you reject the null. You never “prove” H₁ — you show the null is implausible given the data. (Subtle but important: absence of evidence isn’t proof of no effect.)

Three flavors of question (increasingly demanding)

The kind of question sets how hard the design must be:

  • Descriptive — “what / how much / how often?” (What’s the average landing error of PID in wind?) Easiest — just measure carefully.
  • Relational / comparative — “does A relate to / differ from B?” (Does X differ from PID?) Needs a fair comparison + statistics.
  • Causal — “does A cause B?” (Does switching to X cause lower error?) The hardest by far — correlation isn’t causation, and establishing cause demands controlled experiments or careful causal-inference methods to rule out confounds. Be honest about whether your design supports a causal claim or only a correlational one; over-claiming causation from correlation is the classic sin.

When you don’t have a hypothesis (and that’s fine)

Not all good research is hypothesis-testing:

  • Exploratory research — you’re mapping unknown territory (“what happens if we scale this up? what patterns are in this data?”). The output is new questions/hypotheses, not a confirmed one. Legitimate — but don’t dress it up as confirmatory afterward (HARKing — inventing the hypothesis after seeing the results — is a real integrity problem).
  • Systems / engineering research — the question is often “can we build a system that does X, and how well?” The hypothesis is implicit (“this design achieves the goal”); the evaluation is the proof.

The honesty move: know and state whether you’re doing confirmatory (testing a pre-stated hypothesis) or exploratory (discovering) work. Pre-registration — publicly stating your hypothesis and analysis before collecting data — is the strongest way to keep the two honest, increasingly expected in empirical fields.

A good research question, checklist

  • Specific — one clear thing, not a bundle.
  • Measurable — every key term operationalized into a metric.
  • Answerable in scope — finishable with your time/resources (FINER feasibility).
  • Connected to the gap — answering it fills the hole your literature review found.
  • Honest about type — descriptive/relational/causal, confirmatory/exploratory.

Key insight

A fuzzy topic can’t be researched — you must grind it into a sharp, measurable question, and (for confirmatory work) a falsifiable hypothesis that some result could prove wrong. The core skill is operationalizing: turning vague words like “better” and “robust” into named metrics, then framing the study as “change the independent variable, measure the dependent one, hold controls constant, and hunt the confounds that could fool you.” Match the design’s rigor to the question type (descriptive < relational < causal, the hardest), and stay honest about whether you’re confirming a pre-stated hypothesis or exploring — pre-registration is what keeps that line clean.

↑ Back to Research & Paper Writing