PPO on Cartpole

what worked best:

• a low clipping (0.12)
• a low policy estimator number of parameters (80 seems to be enough)
• a policy estimator learning rate not too high (above 0.009 is too high, 0.008 seemed just right)
• not too many layers on the policy estimator
• memory buffer size: not too much. 1000, as I found on the internet, seems just great.
• more epochs is better

The parameters of the function approximator seem to not matter that much, so I'd better not make it too much heavy on computation.