Wiki Home
Reinforcement learning
Definition
Reinforcement learning (RL) trains an agent to choose actions that maximize a cumulative reward returned by an environment, rather than learning from labeled examples. The learning signal is evaluative (how good was the action taken) instead of instructive (what the correct output was). Policy-gradient methods parameterize a stochastic policy, a distribution over actions, and adjust its parameters in the direction that increases expected reward (Sutton and Barto, 1998).
The multi-armed bandit is the simplest case: a single decision with no state, where the agent balances exploiting actions known to pay well against exploring uncertain ones. The gradient-bandit algorithm maintains a numerical preference per action, converts preferences to sampling probabilities with a softmax, and nudges each preference by the received reward measured relative to a running baseline. RL is the natural framing whenever the available feedback is a reward for actions the learner chooses, not a target it is told to reproduce.
Papers that use this concept
- Knockoff Nets: Stealing Functionality of Black-Box Models — its adaptive query strategy is a policy, learned with a gradient-bandit update, that selects which images to send to the victim to improve query sample-efficiency.
See also
References
- Sutton, R.S. and Barto, A.G. “Introduction to Reinforcement Learning.” MIT Press, 1998.