Generalized Risk-Aversion in Stochastic Multi-Armed Bandits
This addresses a theoretical limitation in bandit algorithms for risk-averse decision-making, but it appears incremental as it extends existing frameworks without broad empirical validation.
The paper tackles the problem of minimizing regret in stochastic multi-armed bandits by using a general function of mean and variance as the measure of goodness, rather than just the mean, and characterizes conditions for learnability while showing that sublinear regret is not always achievable.
We consider the problem of minimizing the regret in stochastic multi-armed bandit, when the measure of goodness of an arm is not the mean return, but some general function of the mean and the variance.We characterize the conditions under which learning is possible and present examples for which no natural algorithm can achieve sublinear regret.