Guaranteed Linear Convergence of Projected Policy Gradient for Reinforcement Learning

Projected policy gradient (PPG) under the simplex parameterization exhibits global linear convergence for any constant step size, overcoming the sublinear convergence limitations of previous analyses.