Duality: A New Approach to Reinforcement Learning

The environment and its underlying dynamics of all the reinforcement learning problems are typically abstracted as a Markov decision process (MDP). Because MDPs are useful for studying optimisation problems solved via dynamic programming and reinforcement learning. Today, most of the effective existing reinforcement learning (RL) algorithms are rooted in this dynamic programming paradigm. An alternative paradigm for RL is based on linear programming (LP). The researchers at Google, try to generalise this LP approach and have tried to demonstrate how relevant it is to RL. A Brief Of Overview Of Duality Duality or the duality principle is associated with the optimisation theory, which posits that optimisation problems can be perceived as the primal problem or the dual problem.