The Neural Networks of Slow Learning

For the last few days I have been coding up a neural network for my dancing robots AI so that it can learn how to properly decide when to improvise. It isn't really ready yet, but right now it plays at least as good as the current naive approach to the problem (which is to never improvise no matter what.)

I have been training it for a few hours (which should be enough for a simple neural network at least, but it doesn't seem to be catching on very fast (probably I am not doing a good job of coming up with training data)).

I was figuring that the problem was a difficult one (I didn't want to figure it out personally), but I didn't think that it would be this hard. I will continue to fool around with it a bit, but probably not too much. Perhaps the other stupid parts of the AI will fall to neural networks more easily.

Also, I played some more games of Coin Age.

Comments

Post a Comment

Popular posts from this blog

More Automatown and getting ready for SaltCon

Image
Alison and I played Automatown with many of the discussed changes last night. We liked some of them, and others we did not. The game ended up ending in 6 rounds, and it took about 80 minutes. The later rounds were longer than the first few rounds. We also were being distracted by kids, so perhaps we could have gone faster. Alison ended up with 7 workers and 9 points while I had 9 workers and 28 points (more about that later). Having fewer types of resources was very good. It made the economy less confusing, and setup and cleanup much faster. The mats worked well, and during the game, we thought of more things to put on them - (the three starting workers can be added here, as well as a reminder of what level is what in terms of dots.) Here is the mat that we played with: We started with 3 scrap, which was not enough given that we were also playing with the rule that made upgrades cost a scrap. In fact, that rule made us strapped for scrap the whole game. We decided that m...
Read more

Probably joining the BGDG

Image
I have started on some artwork for witch hunt . It still is not complete by any means, but it is closer to what I want than the white paper with black writing that I used to have. The cards were designed by me, and the art was drawn by Kit Cox. I found it on the Discworld Wiki . I am currently tracking down Kit and asking him permission to use it, but I have not succeeded so far in contacting him. (He is no longer active on the wiki, and the last time he posted there was 6 years ago). I think that I have found him on Facebook, but it remains to be seen weather it really is him or not. I am trying to come up with ways to speed up the game, and I think that lowering the number of testimony cards will do that. Perhaps that will be enough, but perhaps I will have to do more. I am being quoted on another Kickstarter campaign (this time by tasty minstrel games, so a bigger campaign than last time). I have not gotten as many games of Harbour in as I would have liked recently. Wit...
Read more

It's a Doge eat Doge world

Image
I have been preparing for my presentations for the Ion award. I have the beginnings pretty good, but need to practice the sample turns still. I found out that the Sandy board game design group is probably going away, since no one wants to pay a monthly fee for it. It was useful to me when I couldn't make it to the BGDG, but now that my Tuesdays are free, I don't really see the need for two groups anymore. I'm also starting to work on a new game named "Doge city". It is based on the computer game agar.io. Alison and I played a game today, and it mostly works. Perhaps I'll post it some time (after it works and has written rules) I had a dream about a game where the players each had dinosaur species that they were evolving. I woke up and wrote it down and it almost seems playable already. Some time I should make it and see if it is. I also still haven't made my new "giant robots fight alien war machines" game yet. Apparently the mo...
Read more