Convergence of Momentum-based Distributed Stochastic Approximation with RL Applications
Ankur Naskar, Gugan Thoppe · 2023
We develop a novel proof strategy for deriving almost sure convergence of momentum-based distributed stochas-tic approximation (DSA) schemes. Popular momentum-based schemes such as Polyak's heavy-ball and Nesterov's Accelerated SGD can be analyzed using our template. Our technique enables us to do away with three restrictive assumptions of existing approaches. One, we do not need the communication matrix to be doubly stochastic. Two, we do not need the noise to be uniformly bounded. Lastly, our approach can handle cases where there are multiple or non-point attractors. As an application, we use our technique to derive convergence for momentum-based extensions of the multi-agent TD(O) algorithm, where the above restrictive assumptions do not hold.