Exploration–exploitation tradeoff
Also: exploration vs. exploitation, explore-exploit, ε-greedy, epsilon-greedy
An agent that only exploits can get stuck with a mediocre action it happened to try first; one that only explores never cashes in. Strategies like ε-greedy, optimistic initialization, UCB and entropy bonuses balance the two, usually exploring less over time.
Example
With ε-greedy at ε = 0.1, the agent takes a random action 10% of the time and its best-known action otherwise; ε is often decayed from 1.0 toward 0.05 during training.