Skip to content

Retrain the trader: trader-evade-e1 is overfitted to one opponent #1

Description

@cgreening

trader-evade-e1 was trained against pirate-attack-e1 alone, and it shows:

opponent deaths
pirate-attack-e1 (its training partner) 0%
the scripted baseline 18%

Worse against the dumb opponent than the clever one — the signature of
specialising into a single pursuit curve rather than learning to evade.

It is also not wired into the game: brains.ts has no trader-evade import
at all, and an armed trader flies the defence policy (jameson-defend-g1). So
this is not "swap a better brain in", it is "train one worth wiring".

What to do: rotate the opponent. --pool is refused for evade (it
consumes the trader slot, which is the genome under test — see run 18), so the
route is repeated --opponent runs against a mix of pirate-attack-g3,
pirate-attack-r2 and the scripted AI, or a pirate-side pool added to
evolve.ts.

Done when: deaths are comparable across all three opponents rather than
near-zero against one, on held-out seeds.

Context: docs/TRAINING-LOG.md run 18 and its postscript.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions