Probably my favourite phone game (also one of my favourite card games, especially as a 2-player).
I often whinge that a game lacks a single-player option...something that often requires an AI/bot-player. As a non-AI sort of person, I sort of assumed that these would be (while not simple), reasonable standard off-the-shelf utilities. This is a blog post from Temple Games, who adapted Race For the Galaxy into a phone app, and it was a real eye opener for me, and will probably make me complain slightly less (though only slightly) next time a game comes out that either doesn't have bot support, or has something with the intelligence of Battlefield 1942. It does also give me more respect for always shipping Unreal Tournament with decent bots, as it's clearly a non-trivial task.
One thing that really struck me was the size of the AI logic in code terms...a quarter of the full download (and the game has all the cards digitally represented...I normally assume 95% of any download is graphics and sounds, not a massive chunk of logic). They also go over some of the improvements they can make based on the findings of the Alpha Go Zero Project.
Comments
Just got round to reading this through. I think these sorts of games are ideal for AI because there is a statistically best way to play (like Go or chess). The download on iOS is 88MB, a quarter is 22MB. They ship 24 neural networks, which makes 900KB of floating point numbers per network (the code for the network is only a few loops). I imagine they use "single" number precision (float), which means a number sits in a single byte. I say that because mathematics is twice as slow on double byte numbers. So that's 900,000 numbers per network. That sounds about right because the input layer is 800 wide. We'll find out how many neurons they have in total.
Cracking read, thanks mate!