Multi-Armed Bandits for Efficient Lifetime Estimation in MPSoC design
Led by recent MEng graduate Calvin Ma (now with GlobalFoundries), the objective of this work is to accelerate system lifetime estimation in the context of design space exploration by adaptively allocating estimation effort. Monte Carlo Simulation (MCS) samples designs uniformly, expending the same estimation effort regardless of the likely quality of a design. This wastes computational effort on designs that are either obviously good or obviously bad. Multi-armed bandits refer to a class of algorithms that designed to select among alternatives based on an estimate of the quality of each. As more samples are taken, some alternatives become obviously worse than others, and need not be evaluated further. This can be leveraged in design space exploration, where the objective is to differentiate between designs based on lifetime relatively, rather than establish lifetime exactly. This work will be presented at the 2017 Design, Automation Test in Europe Conference Exhibition (DATE).