Regret
Regret is the central measure of performance for online learning algorithms, defined as the difference between the cumulative reward of the optimal policy and that of the actual policy. In bandit problems, minimizing regret is the core objective of algorithm design, and logarithmic regret is the optimal guarantee.
Timeline (last 90 days)
Algorithms such as online gradient descent can guarantee O(√T) regret under convex or approximately convex assumptions.
At the end of life, people hardly ever bring up regrets; that emotion has already disappeared.
Regret in multi-armed bandits is defined as T times the expected reward of the optimal arm after T rounds of play minus the total expected reward obtained by the actual policy.
The regret R_T is defined as the cumulative sum of the differences between the global optimum and the function values at the algorithm's query points over T rounds of interaction, i.e., R_T = Σ[f(x*) - f(x_t)].
All Facts (4)
Algorithms such as online gradient descent can guarantee O(√T) regret under convex or approximately convex assumptions.
50%UnverifiedAt the end of life, people hardly ever bring up regrets; that emotion has already disappeared.
50%UnverifiedRegret in multi-armed bandits is defined as T times the expected reward of the optimal arm after T rounds of play minus the total expected reward obtained by the actual policy.
50%UnverifiedThe regret R_T is defined as the cumulative sum of the differences between the global optimum and the function values at the algorithm's query points over T rounds of interaction, i.e., R_T = Σ[f(x*) - f(x_t)].
50%