Bandit Feedback
A feedback setting in reinforcement learning in which the algorithm can only observe the loss of the action actually executed and cannot learn the losses of other actions, increasing the difficulty of learning.
Timeline (last 90 days)
Future directions include relaxing the oblivious assumption, studying more general privacy models (such as local differential privacy), and extending the technique to other online learning problems (such as bandit feedback).
In online episodic tabular MDPs, the loss function is adversarial, the environment can choose an arbitrary loss at each round, and the feedback is bandit-style, i.e., only the loss of the chosen action can be observed.
All Facts (2)
In online episodic tabular MDPs, the loss function is adversarial, the environment can choose an arbitrary loss at each round, and the feedback is bandit-style, i.e., only the loss of the chosen action can be observed.
50%UnverifiedFuture directions include relaxing the oblivious assumption, studying more general privacy models (such as local differential privacy), and extending the technique to other online learning problems (such as bandit feedback).
50%