Tscmoo terran apparent neural network output

I was watching the new Tscmoo terran with its reputed neural networks.

Hmm, what are those red and blue dots?

detail of apparent neural network output

I read that as the output of the neural network. The dot diagram is incomprehensible unless we know about the network layout. The text is the interpretation; it looks like strategy instructions or hints to the rest of the program. I timed a couple of updates and found them 15 seconds apart, which fits with strategy information.

I can’t tell what the details mean. How can the army composition be tank-vulture if you open with two starports (see those wraiths on the screen)? Is that a prediction for the opponent, maybe? What does “support_wraiths” mean, since I didn’t notice the wraiths seeming to support or be supported by anything?

Trackbacks

No Trackbacks

Comments

tscmoo on Friday, July 29. 2016:

It uses a Recurrent Neural Network architecture called Long short-term memory.
The dots are the values of the memory state and aren't really comprehensible, though it might be possible to associate some of them with game events.
The text below is indeed the output, which is as you guessed instructions/hints, primarily to decide what to build. The outputs are used by a bunch of heuristics and some have unintuitive meanings partly due to the fact that I quickly made some outputs to test and then once I had some data in that format, I couldn't add more inputs or outputs without converting the existing data or starting over. For instance there is a two_barracks flag that will cause it to open with 2 rax, but if it is still set past the early game, it will start making medics and maybe firebats along with more barracks.
In your example, support_wraith was supposed to be a flag to have it make just a few wraiths to "support" the main army, but I messed it up and it ended up just meaning to make as many wraiths as possible off our current starports. army_comp_tank_vulture is just one of the army composition flags, one of which has to be set, but since those flags have low priority, it is instead spending all its resources on wraiths (and some marines, it looks like). If it gets some more economy it will start making tanks and vultures.
My focus is mostly on getting it to respond to previous games against each opponent by opening with the correct build.

Jay Scott on Friday, July 29. 2016:

Cool info, thanks! For everyone else, the original paper on long short-term memory: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.13.634

And this page includes background and a short explanation of what it’s good for: http://people.idsia.ch/~juergen/rnn.html

Hannes Bredberg on Tuesday, June 27. 2017:

I can also recommend this source for learning about LTSM: https://colah.github.io/posts/2015-08-Understanding-LSTMs/

Add Comment

Name*

Homepage

Comment*

In reply to

E-Mail addresses will not be displayed and will only be used for E-Mail notifications.

To prevent automated Bots from commentspamming, please enter the string you see in the image below in the appropriate input box. Your comment will only be submitted if the strings match. Please ensure that your browser supports and accepts cookies, or your comment cannot be verified correctly.
CAPTCHA