pytyche.analysis.TruthComparison

class pytyche.analysis.TruthComparison(cate_rmse, policy_accuracy, rpv_policy, rpv_uniform, rpv_oracle, oracle_gap_rpv)[source]

Bases: object

Sim-mode evaluation metrics. Tight — no optional fields.

Produced by posterior.evaluate_against_truth(tree, truth).

Level: experiment (simulation only).

cate_rmse

RMSE of the posterior-mean contrasts against the truth contrasts, pooled over all n × (K − 1) (visitor, contrast) entries.

policy_accuracy

Fraction of visitors whose tree-assigned arm equals the truth best arm (the shared best-arm rule applied to the truth contrast vectors).

rpv_policy

Realized revenue-per-visitor of the tree policy under the truth potential-outcome surfaces (p_a · m_a per arm).

rpv_uniform

Realized RPV of uniform random assignment.

rpv_oracle

Realized RPV of the per-visitor best arm.

oracle_gap_rpv

rpv_oracle − rpv_policy — how much the policy leaves on the table against a truth-knowing oracle.

Parameters:

  • cate_rmse (float)

  • policy_accuracy (float)

  • rpv_policy (float)

  • rpv_uniform (float)

  • rpv_oracle (float)

  • oracle_gap_rpv (float)

__init__(cate_rmse, policy_accuracy, rpv_policy, rpv_uniform, rpv_oracle, oracle_gap_rpv)
Parameters:

  • cate_rmse (float)

  • policy_accuracy (float)

  • rpv_policy (float)

  • rpv_uniform (float)

  • rpv_oracle (float)

  • oracle_gap_rpv (float)

Return type:

None

Methods

__init__(cate_rmse, policy_accuracy, ...)

Attributes

cate_rmse

policy_accuracy

rpv_policy

rpv_uniform

rpv_oracle

oracle_gap_rpv