Choose actions with epsilon-greedy exploration, deterministic greedy tie handling, and controlled random sampling.
