Online distributed optimization with stochastic gradients: high probability bound of regrets

关键词:  Distributed optimization · Online optimization · Stochastic gradient · High probability

Key words:  Distributed optimization · Online optimization · Stochastic gradient · High probability

摘要:

In this paper, the problem of online distributed optimization subject to a convex set is studied via a network of agents. Each agent only has access to a noisy gradient of its own objective function, and can communicate with its neighbors via a network. To handle this problem, an online distributed stochastic mirror descent algorithm is proposed. Existing works on online distributed algorithms involving stochastic gradients only provide the expectation bounds of the regrets. Different from them, we study the high probability bound of the regrets, i.e., the sublinear bound of the regret is characterized by the natural logarithm of the failure probability’s inverse. Under mild assumptions on the graph connectivity, we prove that the dynamic regret grows sublinearly with a high probability if the deviation in the minimizer sequence is sublinear with the square root of the time horizon. Finally, a simulation is provided to demonstrate the effectiveness of our theoretical results.

Abstract:

In this paper, the problem of online distributed optimization subject to a convex set is studied via a network of agents. Each agent only has access to a noisy gradient of its own objective function, and can communicate with its neighbors via a network. To handle this problem, an online distributed stochastic mirror descent algorithm is proposed. Existing works on online distributed algorithms involving stochastic gradients only provide the expectation bounds of the regrets. Different from them, we study the high probability bound of the regrets, i.e., the sublinear bound of the regret is characterized by the natural logarithm of the failure probability’s inverse. Under mild assumptions on the graph connectivity, we prove that the dynamic regret grows sublinearly with a high probability if the deviation in the minimizer sequence is sublinear with the square root of the time horizon. Finally, a simulation is provided to demonstrate the effectiveness of our theoretical results.