Each curve is the mean cumulative regret across repeated runs; the faint envelope is the 20th–80th percentile range. More averaged runs make the estimate smoother. More pulls per run make constant exploration costs visible.

ε-greedy can look strong at short horizons. Its weakness is the fixed random tax: it keeps exploring after the answer is mostly clear. Use 2,000 or 5,000 pulls/run to see that cost; use more averaged runs to reduce simulation noise.