On 7/20/21, Matt Mahoney <[email protected]> wrote:

> The paper describes an experiment in which a neural network was trained to
> play tic tac toe. But instead of describing what was actually done, here is
> a meaningless graph that it produced and a link to the source code.

The paper described the neural network architecture in the previous section.
The graph shows early convergence, but afterwards convergence is very slow
and unstable with respect to the optimal value.  This unstable behavior is also
observed in the "control" test, so it may be due to the policy
gradient algorithm
itself, rather than the symmetric neural network.

> In the pages where we would normally find a description of the algorithm
> and the motivation for doing it this way, we instead find an incoherent
> collection of math and logic tidbits that are irrelevant to the problem at
> hand.

I explained how logic formulas are treated by the symmetric neural network,
and then the tic-tac-toe problem was tested with such an architecture and
with propositional representation of the play moves.  What do you find
incoherent?

> Does the tic tac toe program use a Curry Howard transform or the pair of
> pants from string theory? Did it help? How will this shorten the path to
> AGI?

Curry-Howard provides an interpretation, which is not essential for
designing the symmetric neural network.  But Curry-Howard is essential
to understanding the whole theory of categorical logic... it's at the very
foundation of the theory and cannot be ignored.  So I spent some time
explaining it in the paper, otherwise readers would be lost when trying
to read the texts that are referenced.

The "pair of pants" is just mentioned for general interest... it doesn't
play a role in the current implementation...

Like I said in the introduction, I hope that describing AGI in mathematical
terms will enable more people to discover connections between AGI
and other mathematical structures....

YKY

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tb5526c8a9151713b-M72c3d74a8da43883dcb10751
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to