/* ---- Google Analytics Code Below */

Friday, November 25, 2022

Optimal Decision Making

 Technical, but quite interesting point being made.  Optimality may be a good thing,  but how do I embed it in useful real time decisions?  Notably too the consideration of noise, often a key consideration.  This worth a look. 

Algorithm for Optimal Decision-Making Under Heavy-Tailed Noisy Rewards

Chung-Ang University (South Korea), November 17, 2022

Researchers at South Korea's Chung-Ang University (CAU) and Ulsan Institute of Science and Technology created an algorithm that supports minimum loss under a maximum-loss scenario (minimax optimality) with minimal prior data. The algorithm addresses sub-optimal performance for heavy-tailed rewards by algorithms designed for stochastic multi-armed bandit (MAB) problems. CAU's Kyungjae Lee said the researchers proposed minimax optimal robust upper confidence bound (MR-UCB) and adaptively perturbed exploration (MR-APE) methods. The team obtained gap-dependent and independent upper bounds of the cumulative regret, then assessed their methods via simulations conducted under Pareto and Fréchet noises. The researchers found MR-UCB outperformed other exploration techniques with stronger robustness and a greater number of actions under heavy-tailed noise; MR-UCB and MR-APE also could solve heavy-tailed synthetic and real-world stochastic MAB problems.

Full Article

No comments: