Regret Bound
A measure of algorithm performance in online learning, defined as the difference between the model's cumulative loss and the cumulative loss of the best fixed strategy in hindsight. A sublinearly growing regret bound implies that the average loss is asymptotically optimal.
Timeline (last 90 days)
For the online learning setting, every consistent online learner achieves the optimal regret bound.
The algorithm proposed in the paper achieves, in adversarial tabular MDPs, regret bounds of Õ(√(HS(H+A)T)) under known transitions and Õ(HS√(AT)) under unknown transitions.
The regret bound Õ(HS√(AT)) under unknown transitions in the paper matches the known optimal bound for adversarial tabular MDPs.
All Facts (3)
For the online learning setting, every consistent online learner achieves the optimal regret bound.
50%UnverifiedThe algorithm proposed in the paper achieves, in adversarial tabular MDPs, regret bounds of Õ(√(HS(H+A)T)) under known transitions and Õ(HS√(AT)) under unknown transitions.
50%UnverifiedThe regret bound Õ(HS√(AT)) under unknown transitions in the paper matches the known optimal bound for adversarial tabular MDPs.
50%