exploration strategies

An agent only learns about the actions it actually tries, so if it always grabs whatever looks best right now, it may never discover something far better. Exploration strategies are the family of recipes for deliberately trying uncertain or unfamiliar options, trading a little immediate reward for information that pays off later. They are the practical answer to the exploration–exploitation dilemma that sits at the heart of reinforcement learning.

The strategies range from blind to clever. The simplest just inject randomness — act greedily most of the time but occasionally pick at random. Smarter ones are directed: they steer toward states the agent has rarely seen, where its predictions are poor, or where its uncertainty about value is largest. Some carry an explicit model of what they do not know; others bolt curiosity rewards onto the environment's own. The right choice depends on how sparse the reward signal is and how deeply it is buried.