ε-greedy
A classical exploration strategy that selects a random action for exploration with probability epsilon and selects the currently estimated best action with probability 1 minus epsilon. Its computational cost is constant time, and it is commonly used in multi-armed bandits and reinforcement learning.
Timeline (last 90 days)
E2D.PMSide adopts an ε-greedy strategy in the globally observable setting, achieving the optimal regret bound at extremely low computational cost.
Whether the worst-case performance of ε-greedy and IGW exactly matches the information-theoretic lower bounds, and the regret-order results given in the paper require more follow-up work to verify their tightness on practical problems.
Common exploration approaches include the ε-greedy strategy (randomly exploring with probability ε, otherwise executing the current optimal action) and entropy regularization (encouraging action diversity in policy gradient methods).
引入 ε-greedy 策略并让 ε 随训练进程逐渐衰减，可以平衡探索与利用，提升学习效果
All Facts (4)
E2D.PMSide adopts an ε-greedy strategy in the globally observable setting, achieving the optimal regret bound at extremely low computational cost.
50%UnverifiedWhether the worst-case performance of ε-greedy and IGW exactly matches the information-theoretic lower bounds, and the regret-order results given in the paper require more follow-up work to verify their tightness on practical problems.
50%UnverifiedCommon exploration approaches include the ε-greedy strategy (randomly exploring with probability ε, otherwise executing the current optimal action) and entropy regularization (encouraging action diversity in policy gradient methods).
50%Unverified引入 ε-greedy 策略并让 ε 随训练进程逐渐衰减，可以平衡探索与利用，提升学习效果
50%