This animation shows how feature subsets can be represented as binary masks and searched using random search and genetic algorithms. It covers evaluation budgets, caching repeated subsets, crossover and mutation operators, constraint checking on feature counts, and combining accuracy with feature cost. Nested validation separates model selection from final testing, and repeated runs demonstrate selection stability rather than guaranteed optimality. Useful for first-year AI students learning search-based feature selection and sound validation practice.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# Random feature search needs a budget and an honest evaluation Represent a four-feature subset as a binary mask: 1010 selects features 1 and 3. Random search proposes masks, evaluates them using training-internal validation, and keeps the best result within a fixed evaluation budget. Cache repeated subsets rather than charging them as new discoveries. Next illustrate genetic search: parents 1010 and 0101 cross after the second position to produce 1001 and 0110, followed by a clearly marked bit mutation. If the number of selected features is constrained, check each offspring against that rule. Introduce an optional score combining validation performance and feature cost, and distinguish this design choice from raw accuracy. Show nested validation around the entire selection process, with final test data untouched. Repeated runs reveal selection stability rather than proving causal importance. Use black backgrounds, clean mask cards, and English explanations. Useful for understanding stochastic search without mistaking randomness for a guarantee of optimality. Audience: 一年级人工智能专业学生