Understanding the AI Problem
Free Certified Artificial Intelligence Practitioner practice — 6 questions on Understanding the AI Problem, with explanations. No sign-up. Full 12-question mixed test →
A logistics company wants warehouse robots to learn optimal picking paths through trial-and-error, receiving a reward signal when a task is completed efficiently and a penalty for collisions or wasted movement. No labeled path examples exist. Which ML paradigm BEST fits this problem framing?
The scenario describes an agent interacting with an environment, receiving rewards/penalties, and learning a policy over time with no pre-existing labeled dataset — the defining characteristics of reinforcement learning, a core robotics use case tested on the exam.
A media company has a large corpus of unlabeled news articles and wants to automatically discover underlying themes without any predefined category list. Which approach BEST matches this business requirement?
Because no labels or predefined categories exist and the goal is to discover latent structure in text, this is an unsupervised learning problem — topic modeling is the standard technique for this use case.
A manufacturer is framing a predictive maintenance problem where equipment failures make up only 2% of historical records. Stakeholders want to catch as many true failures as possible while limiting costly false alarms. Which evaluation approach should be established during problem framing?
With a 2% minority class, a naive model predicting 'no failure' every time would score 98% accuracy while catching zero failures. Precision, recall, and F1 explicitly capture the trade-off between missed failures and false alarms that stakeholders care about.
A fraud detection team is framing a classification problem where missing an actual fraud case (false negative) costs the company roughly 10 times more than investigating a false alarm (false positive). During problem framing, how should the classification decision threshold be adjusted relative to the default 0.5?
When false negatives are far more costly than false positives, lowering the decision threshold makes the model flag more cases as 'fraud,' trading some additional false alarms for fewer missed fraud cases — aligning the model's behavior with the stated business cost asymmetry.
During the AI problem framing phase, why is it important to establish a human baseline performance metric before selecting an algorithm?
A human baseline gives stakeholders a concrete reference point to judge whether a proposed ML solution is worth the investment — if the model can't beat (or reasonably match at lower cost) existing human performance, automation may not be justified.
A call center wants to build a transcription system for customer calls in a low-resource language, with only 2 hours of labeled audio available and no existing pretrained model for that language. Which approach is MOST realistic when framing this AI problem given the data constraints?
With extremely limited labeled data, transfer learning from a pretrained multilingual acoustic/language model dramatically reduces the data required, and combining it with rule-based fallback for common phrases is a realistic, exam-tested strategy for low-resource speech problems.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →