The Epsilon-Greedy /UCB ("upper confidence bound") for MAB (Multiarmed-bandit) problem sometime in reinforcement learning (RL)

你是球队教练,现在突然要打一场比赛,手下空降三个球员,场上只能有一个出战,你不知道他们的能力,只能硬着头皮上,如何根据有限的上场时间看出哪个球员厉害,然后多让他上,从而得更多分数?

Epsilon-Greedy

supposed an k arm(slot) and set ε a little number between [0,0.1]

In short, epsilon-greedy means pick the current best option ("greedy") most of the time----(1-ε) + ε/k

but pick a random option with a small probability sometimes for other option-----(k-1)ε/k

often works as well as, or even better than, more sophisticated algorithms such as UCB

for more information about

A/B testing

Thompson sampling

see

https://towardsdatascience.com/solving-multiarmed-bandits-a-comparison-of-epsilon-greedy-and-thompson-sampling-d97167ca9a50

猜你喜欢

转载自www.cnblogs.com/yifan2015/p/12005552.html