Discrete-to-deep reinforcement learning methods

Neural networks are effective function approximators, but hard to train in the reinforcement learning (RL) context mainly because samples correlated. In complex problems, a neural RL approach is often able learn better solution than tabular RL, generally takes longer. This paper proposes two methods, Discrete-to-Deep Supervised Policy Learning (D2D-SPL) and Q-value (D2D-SQL), whose objective acquire generalisability of network at cost nearer that method. Both methods combine supervised (SL) based on idea fast-learning method can generate off-policy data accelerate RL. D2D-SPL uses classifier which then used as controller for problem. D2D-SQL initialise allowed continue using another We demonstrate viability our algorithms with Cartpole, Lunar Lander an aircraft manoeuvring problem, three continuous-space environments low-dimensional state variables. least 38% faster baseline yield policies outperform them.