Distributed Consensus Optimization with Network Delays over Directed Graphs

This release contains the ADD-OPT with delays implementation for a huge dataset of delays and stepsizes. The aim was to analyze the system stability with respect to varying stepsizes and delays and also the critical points beyond which the system is unstable i.e, the system starts diverging in place of converging. The necessary scripts for testing and the respective logs are stored in the repository.

ADD-OPT with delays are simulated successfully. The results obtained here are analyzed and will be summarized as my master thesis. The thesis objectives are met for this release and, the work in coming releases would contribute to future work. :)

Synchronous Distributed Network Optimization Over Directed Graphs Using Row-stocahstic Weights

This release adds two new algorithms to the collection which uses row-stochastic weight matrices to solve the optimization problem in a distributed manner. The row-stochasticity enables the agent to associate weights to incoming neighbors, that way the dependency on knowledge of out-neighbors is eliminated. In simple words, algorithms presented in this release are more realistic and from the convergence plots, it is evident that consensus is achieved faster than previous algorithms. Also, from the residual plots, we can conclude that FROST linearly converges to the optimal solution.

FROST Implementation has been completed considering networks with no delays. In the upcoming release (probably the last release), we will be working ADD-OPT and FRSOT with asynchronous networks. In asynchronous networks, information exchange between agents can be delayed and also graph topologies are time-variant.

Synchnronous Network Consensus Algorithms over Undirected graphs

This release includes three major algorithms (Pushsum, Subgradient-push, and ADDOPT) which are among well-known solutions to distributed optimization over undirected graphs using doubly-stochastic or column-stochastic weight matrices. The column stochasticity is associated with the out-neighbors and it is always known information for any distributed network. This infeasibility drives us forward and to explore the algorithms using row-stochastic weights as they are associated with the in-degree of agents. Hence, the focus for the upcoming release will be on one such popular algorithm known as FROST (Fast Row-Stochastic Optimization with uncoordinated STep-sizes)

Also, with the present release and implementations, the master thesis completes its second milestone.

ADDOPT Consensus Algorithm for Synchronous Networks

This release contains a new consensus algorithm named ADDOPT algorithm for synchronous networks. Earlier release codebase has been added with generic functionalities to work with any cost functions.

Pushsum and Subgradient-push Algorithms for Synchronous networks

This update is the initial commit of the project containing consensus algorithms for directed graphs with a column-stochastic weight matrix and assuming synchronous update of local agent information to its neighbors.