Building a too long string results in a memory error. Using a table to
concat afterward fix this.

Use a more explicit argument name (so called dump) and recurcively load
arguments if necessary.

This can be used by tools like t-SNE

Implement my own logging system as classes to avoid logger dumping

This can be used by tools like t-SNE

Try to use a random network as starting network with real data to see
the impact of the pre-train phase.

These layers make the network not learn anything...

Remove this layer which move output's mean far below zero (around -3)
and seems to kick the reward (ranging between -1 and 1)

Big reward range doesn't converge as well as expected, try with a
narrower one.

Initial values of the network seem to give a value of -3 to all
Q-states. Rewards of +/-1 are to long to make the network learn, so try
with wider rewards (+/-10)