tauopt.engine.policy.default#

The open default policy. The open default must be STRONG: the closed router is a delta on a strong baseline, not the difference between working and broken.

Routing: space-filling bootstrap until max(4, 2 * dim) observations, then GP + LogEI with constant-liar handling of open suggestions – or qLogNEHVI when >= 2 metrics are optimized.

FALLBACK_LADDER is the never-fail order, executed by tauopt.engine.ask – gp_ei timeout/error -> ucb_cl -> sobol -> random. Multimetric states run MOO_FALLBACK_LADDER (ucb_cl is single-objective, so the MOO ladder skips it). The final rung is stdlib-only and cannot fail; a fallback is recorded in provenance, never hidden.

Module Contents#

Classes#

DefaultPolicyV1

default@v1 – the tag since 0.3.0 (dedup default-on): the tag rolls ONLY on a behavior-moving change, and dedup-on changes what default policy asks emit on collision-bearing trajectories. Routing itself is unchanged from v0.

Data#

FALLBACK_LADDER

MOO_FALLBACK_LADDER

API#

tauopt.engine.policy.default.FALLBACK_LADDER: tuple[str, ...]#

(‘gp_ei’, ‘ucb_cl’, ‘sobol’, ‘random’)

tauopt.engine.policy.default.MOO_FALLBACK_LADDER: tuple[str, ...]#

(‘gp_nehvi’, ‘sobol’, ‘random’)

class tauopt.engine.policy.default.DefaultPolicyV1#

default@v1 – the tag since 0.3.0 (dedup default-on): the tag rolls ONLY on a behavior-moving change, and dedup-on changes what default policy asks emit on collision-bearing trajectories. Routing itself is unchanged from v0.

name#

‘default’

version#

‘v1’

route(state: tauopt.engine.state.ExperimentState) → tauopt.engine.config.OptimizerConfig#

Route this ask: sobol until max(4, 2*dim) non-failed observations exist, then gp_ei, or gp_nehvi when the state is multimetric. Failed observations are not evidence; open suggestions do not advance the bootstrap.