Upgrade to flow-0.5.0.dev & remove examples(#15)

Update 0x01/README.latex plus additions (#20)

Delete manually edited osm and net files (#17)

Reduce train, evaluate and re-train cycle (#14)

ISSUE:
------

Reduce the hyper parameter search cycle

FIX:
---

The longest task during training was the generation of the emissions
file which always grows do the decimal of the GBs and took a long time
to generate. The shift in scripts was always allow the info json (for
aggregate performance metrics) and the pickle (for recovering the
training models) to be generated by default. While enabling emissions'
file generation only on evaluation.

Update envs' reward and default distrib. (#14)

Tune speed and count categories, fix tests (#14)

Refactor baseline scripts into generic ilu (#13)

Stabilize baseline intersection scenario (#13)

Add flow and queue features to TLSAgent

	STATES:
	-------
	States were described by discretization of two variables
mean speeds and mean number of the vehicles trafficking over the
intersection scope (both incoming and outgoing sections).
 	Now two new variables have been added the mean throughput
(flow) and queue as it appears such variables are more common in
traffic light signal control literature. The flow is computed
only on the outgoing sections (everything that's exiting must have
entered at some point, while queues are computed only over incoming
section edges (outgoing edges make the incoming edges of neighbours)
	In addition to two new variables the variables are now
composable i.e before only mean speeds and mean number were allowed
now we may run experiments with either flow or queue but with both
together.

	ACTIONS:
	--------
	The choice criteria which only counted with epsilon greedy
now was expanded to incorporate the upper bound confidence interval.
see section 2.7 of [1]

    	REFERENCES
    	----------
    	[1] Sutton et Barto, Reinforcement Learning 2nd Ed 2018

Show statistics plot dynamically while training