Closed migarstka closed 6 years ago
The solution follows from Boyd - Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers s.3.4.3. Over-relaxation
In the z- and y-updates the quantity Ax(k+1) can be replaced with alpha*Ax(k+1) - (1-alpha)(Bz(k)-c).
For my ADMM equation the constraint is x=s, hence I do the following:
xNew = from linear system
xRelax = α*xNew+(1-α)*sPrev
#relaxed s and lambda update
sNew = Projections.sdcone( xRelax + (1/σ)*λPrev,r)
λNew = λPrev + σ*(xRelax - sNew)
Wrong implementation of relaxation seems to be the major bug causing the convergence problems!