sketch_epsilon_greedy.nx

sketch_epsilon_greedy.nx -- Epsilon-Greedy bandit (Sutton-Barto, classic). Simpler alternative to UCB1. At each pull: with probability epsilon: pull a UNIFORM-RANDOM arm with probability 1-epsilon: pull arm with highest empirical mean Three modes via the epsilon_mode field: NX_EG_MODE_FIXED -- epsilon constant (param: epsilon_ppm) NX_EG_MODE_DECAY_LINEAR -- epsilon(t) = epsilon_0 / (t+1) NX_EG_MODE_DECAY_LOG -- epsilon(t) = epsilon_0 / log(t+e) COMPLEMENTS UCB1: - UCB1: deterministic confidence-bound; near-optimal regret; requires sqrt + log per arm per pull - epsilon-Greedy: stochastic; simpler; tunable explore rate; preferred when log/sqrt are expensive (embedded) or when bounded exploration budget needed Together they cover the "classic bandit" textbook coverage. LCG-based deterministic RNG so two runs with same seed produce byte-identical action sequences. LOSSLESS-LANGUAGE DISCIPLINE: nx_eg_query reports the running mean of an arm with stderr ~ 1/sqrt(count_i).