ddpg(Deep Deterministic Policy Gradient) algorithm

tensorflow >= 1.4

Consider :
Critic : (state,action) -> q value(scalar) goodness of action
Actor : state -> action

Deterministic policy :
Action Value function :
Critic network is q function(action-value function),giving the goodness of actor. Training critic network close to the fact (real reward), and supply gradient for actor update.Critic judge the actor is real(close the fact ) or fake(depart from the fact). Policy network update action policy by replay and giving the best action. Actor always try to get the biggest reward.