A stopping criterion determines when an experiment or sequential decision process terminates. In metric-driven experimentation, it balances further exploration against the cost of continuing to test candidate system changes. section 3.1.3 section 3.3.3
For decaying-epsilon greedy selection, epsilon falls as measurements accumulate. The algorithm stops when epsilon is below a chosen threshold, such as 0.01, and can then accept the version with the higher mean CTR. Lower thresholds extend the run; moderate threshold changes have little reported effect on CTR. section 3.1.3
Decaying-epsilon configuration
With bm_max = 0.01, practical significance = 0.001, and c = 5, the implementation computes epsilon_0 = 2c(bm_max/prac_sig)^2 and stops when epsilon < 0.01. The parameter c controls pure-exploration time: values that are too low risk premature exploitation, while values that are too high increase runtime without benefit. section 3.1.3
Thompson sampling uses a probability-based stopping rule built around pbest(arm), while also enforcing enough observations to resolve practical differences. With practical significance 0.001 and a smallest sum difference of one, the minimum is 1/0.001 = 1,000 samples per arm. section 3.3.3
