Interesting vid! The ‘Theory of Everything’, and physical Unified Field Theory being only a part of it ? Wow – this guy has really got a vision! And all this uncovered by AI algorithms. Forgive me if I’m a bit doubtful about this.
Google brought his company for $400 Million, maybe the saw something of interest.

Back to our ‘simple’ FX business. After having learnt a bit more about the general approach, I might see those 52% success rate in a different light now, and I wonder what would happen if we modified what the speaker called the ‘Environment’ in his slides:
a) offer Deep Q a longer data history to learn from
Yes you can feed the network more data anytime frame any market, I feed it with 66,000 rows of data, but you can add more if you wish

b) offer it a broader data universe, say all OHLC+V of all crosses of EUR, CHF, GBP, USD, CAD, AUD, NZD, JPY to learn from
Yes

c) offer it a dynamic model of pair specific trading cost
and then let it learn.
Yes, this also fine and good idea

Learn for how long?
Data depend few minutes to a few hours

Which computing resources (CPU, RAM) provided?
The higher the better, large network models can from the use of super computer this can be hired on a time basis from Amazon,

Question above all: What exactly is the ‘Goal’ we’re setting up for the learning process?
Goal is simply a reward I choose maximum pips per trade an example you can use percentage correct as a Goal/Reward if you wish

Secondary question: is it wise to add additional time series, say indicator data, or would this compromise the uniqueness of the approach?
Yes this is fine and encouraged

Anyway, I’m still interested in some conceptual papers, rather than conference speaks. Sorry for being so oldfashioned!

Yes paper available at http://arxiv.org/pdf/1312.5602v1.pdf

And I would be interested in some modeling details about that ‘marketcheck’ link you provided in your first post.

Currently working on site will be available in the future