Development of model free flight control system using deep deterministic policy gradient (DDPG)

Developing a flight control system for a complete 6 degree-of-freedom for an air vehicle remains a huge task that requires time and effort to gather all the necessary data. This thesis proposes the use of reinforcement learning to develop a policy for a flight control system of an air vehicle. This method is designed to be independent of a model but it does require a set of samples for the reinforcement learning agent to learn from. A novel reinforcement learning method called Deep Deterministic Policy Gradient (DDPG) is applied to counter the problem with large and continuous space in a flight control. However, applying the DDPG for multiple action is often difficult. Too many possibilities can hinder the reinforcement learning agent from converging its learning process. This thesis proposes a learning strategy that helps shape the way the learning agent learns with multiple actions. It also shows that the final policy for flight control can be extracted and applied immediately for a flight control system.