AI keeps spamming Basalt Monolith untap ability and other AI problems

[TakeHeedStruggler]

The Ai keeps spamming the untap ability of Basalt Monolith, resulting in stalled games

[spjspj]

This has been an issue for at least 7+ years :) The solution is to not give the AI complicated decks with cards like basalt monolith.

[JayDi85]

Xmage has some logic to protect from AI repeated action. Maybe it can be improved.

[jeffwadsworth]

I had the GPT models look for alternative ways to validate repeat actions but nothing worthwhile came of it. So, my advice is to wait for a GPT or Open-Assistant AI player API, or just don't let the current AI have the card. GPT-4 can play a board state incredibly well.

[JayDi85]

[[Basalt Monolith]]

[github-actions[bot]]

Basalt Monolith - (Gatherer) (Scryfall) (EDHREC)

{3} Artifact Basalt Monolith doesn't untap during your untap step. {T}: Add {C}{C}{C}. {3}: Untap Basalt Monolith.

[JayDi85]

Added reproduceable AI test in e3de777bd564e7685bd1376641ad61ec86666d1c.

Manual test (AI will randomly activates an untap ability before creature cast):

[test Basalt Monolith]
battlefield:Computer:Basalt Monolith:2
hand:Computer:Aesthir Glider:1

[JayDi85]

Basalt Monolith catches three different problems (combination of problems):

1. [x] 1 - Wrong search for repeated actions in calculated actions chain - so it try to use it multiple times. Can be fixed. Can be related to 3 (see below);
2. [x] 2 - AI randomly (50%) ignore actions chain result - it's randomly stop possible actions chain. It's look like "oh, I forget about that combo or target combination". Must be disabled on debug. I don't think it has bad effect on gameplay - but it helps to choose different targets/abilities on each activation;
3. [x] 3 - AI don't have tools to find "do nothing" abilities (abilities that has same game score). Need more research. I think it can be related to stack usage -- cause it generates two actions (first activating with minus score to put card from hand to stack or tap, second pass/resolve with plus score to put card to other zone or untap). Possible solution: use stack zone to calculate game score.

[JayDi85]

4. [x] 4 - AI code uses weird best node (game simulation) searching with direct data replaces from one game to another (buggy optimization). So it lost all best chain actions if it was second+ (e.g. if it found best node second+ time then it lost all child nodes). Must clear childs after whole level finish.
5. [x] 5 - AI must find better actions chain (combo of abilities to activate). But for some reason it doesn’t use it after calc. In many real games/tests AI will use only first calculated action (still from a better chain) and will ignore left actions (but must continue to execute it). Maybe related to wrong game hash, data lost from 4 (see above), or it’s made by design.

[JayDi85]

• [x] 6 - Wrong best node logic and game hash usage - on multiple actions chains AI can wrongly use outdated/same targets from another actions chain and fizzle that spells (example: use 2 bolts on 1 creature instead 2 creatures). Actual only for big games with many possible actions (many spells and targets on battlefield). Need more research.

[JayDi85]

• [ ] 7 - Outdated support of "search your library for" effects. AI can run simulations for best card choice, but it support abilities with extends SearchEffect only. Many cards uses custom OneShotEffect classes instead. Need to reseach -- is it workable at all -- and delete or improve support. Maybe it can support any card choices.

Example:

[JayDi85]

• [ ] 8 - AI stuck on take control of other player (maybe related to ai vs human games, maybe for ai vs ai too). Related issues: #12275, https://github.com/magefree/mage/issues/2088#issuecomment-854221220

[JayDi85]

• [ ] 9 - unsupported “choose modes” simulations for triggered abilities like ETB (it select first available mode)