A few improvements incoming:
- I now have a jovian scorer. In theory I'd like this to train purely on the nets but it was playing multipliers alone because it learnt 'they're good cards'. So just added a block of code to manually score this in place of learnt information. Jovian points with multipliers seem a bit too deep for this bot to get and i don't have the money to train a really large deep net so things like this need some helper code.
- It should now counter draft. I never had the nets wired up to "what cards would the other player want". They are there and training now and should deny based on what cards it would want if you."
- Milestone risks tightened up. It see's ending turn 2 cities/greneries behind if opponent goes first next turn as risky.
- Awards risks fixed. Eg. buying mining when starting as IC clearly wrong. I didn't have anything estimating buy down of current resources so it just saw "20 ahead" with no thought to generations left. Now in place.
- Placement grid now matches the AI. Expect improved placements. There was a BGA<>internal AI mismatch before
- Draft picks and buys use the same neural net. It diverged and i had two nets for what should be the same thing. The draft picks was using one from many generations ago.
- If there's cards with requirements some weighting is now given to things that help meet the requirements. Eg. Science tags for mass converter or oceans for Kelp.
The above should bump the AI difficulty reasonably high. Have fun and good luck!
