Tune AI brains for competitive self-play #11
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Current bench (500 games, 5 presets): conservative 291 wins, balanced 187, lender 22, hawk 0. Hawk averages catastrophic equity (its deposit-rate-first archetype pays for funding it cannot deploy). Goal: all four brains win non-trivially in the round robin. Iterate by running the bench and adjusting the Lisp, the way Robert described: play them against each other and previous versions.
Brains are now four structurally different playbooks, not one brain with four dials: empire (deposit franchise, capacity play), mill (leveraged loan volume), vault (treasury + vulture on failed banks), adapt (patient opportunist). Each reacts to different market conditions and each owns different presets in the 500-game bench (current: vault 211 wins, mill 193, adapt 94 incl. 88 in hot-rates, empire 2). Empire's deposit-funded model is structurally weak against equity-funded vault lending in this economy; the churn pool does not reward paying for deposits enough. That is the open economy question, tracked here rather than papered over.
Round three after Robert's review: (1) API is Lisp-idiomatic hyphenated names, no dot notation; (2) dead hawk/lender/conservative/balanced files removed from the repo; (3) every threshold in all four brains is now computed from config, default risk from the die rank, bond prices from the configured face, cash gates from loan sizes, and brains refuse loans when risk clears the cap. Bench expanded to 13 presets including $100 and $21,000 face values and d10 dice. 1300 games: mill 919, vault 219, adapt 101, empire 61, each brain owning different presets. Empire's deposit-funded model is still the weakest line and remains the open economy question.
Closing. The four brains are structurally distinct and each owns different markets in the 13-preset, 1,300-game bench (latest: mill 864, vault 274, adapt 116, empire 46). That clears the 'not bad' bar set on 9/28. Further tuning happens when a new brain or mechanic lands, not as an open-ended issue.