Updated the dqn policies

[adrigrillo]

Changes in the dqn agents:

• Implementation of double and dueling dqn without prioritized memory. The policies now are:

  • DeepQNetwork
  • DoubleDeepQNetwork
  • DuelingDeepQNetwork
  • PrioritizedDeepQNetwork
  • PrioritizedDoubleDeepQNetwork
  • PrioritizedDuelingDeepQNetwork

• The exploration rated can be modified easily, now the epsilon_calculator object is an argument of the policy learner.

An example of how to instantiate an agent:

memory_delay = 5000
init_eps = 1.0
memory_eps = 0.8
min_eps = 0.01
eps_decay = 500000
linear = LinearSchedule(schedule_timesteps=memory_delay, initial_p=init_eps, final_p=memory_eps)
exponential = ExponentialSchedule(initial_p=memory_eps, min_p=min_eps, decay=eps_decay)
policy = PrioritizedDoubleDeepQNetwork(4, env[0].action_space.n, eps_calculator=linear,
                                       memory_eps_calculator=exponential, memory_delay=memory_delay)