Tune hyperparameters of DQN and fix bug

[liyc-ai]

On CartPole-v1, DQN with the original hyperparameters can achieve a return up to ~200. But with the tuned hyperparameters, it can get 500 and be much more stable. I mainly change the hyperparameters in the following ways:

1. increase the learning rate to $3e^{-4}$, the best learning rate of Adam;
2. decrease the target update frequency, making the target not too different from the current network;
3. increase the buffer size, to store much more samples to improve the evaluation accuracy and sample efficiency;
4. change the epsilon schedule, including decreasing the epsilon_begin and the epsilon_steps. In fact, CartPole does not need such aggressive exploration. And I guess the reason why the original DQN performs not well is the totally random actions in the buffer for long timesteps, indeed caused by the epsilon schedule and small buffer.

[mansicer]

These changes are beneficial. Thanks!