Audit a logged policy with importance weights

Advanced65 min

Implement policy_report(events, max_weight=None), returning ordinary IPS, self-normalized IPS, effective sample size, and event count for logged reward observations.

Your task

Each event is a dict with finite numeric reward, propensity, and target_probability. Require 0 < propensity <= 1 and 0 <= target_probability <= 1.
Compute weight = target_probability / propensity. If max_weight is supplied, require a finite positive value and cap each weight.
Return {"ips": weighted_reward_sum / n, "snips": weighted_reward_sum / weight_sum, "ess": weight_sum squared / sum of squared weights, "n": n}.
For empty input return ips=None, snips=None, ess=0.0, n=0. For all-zero weights return ips=0.0, snips=None, ess=0.0.
Reject malformed or nonfinite numeric values with ValueError. The estimator cannot establish support for unobserved actions; explain that limitation.

Examples

EXAMPLE 1

InputRewards [1, 0], logged probabilities [.5, .5], target probabilities [1, 0]

Output{"ips": 1.0, "snips": 1.0, "ess": 1.0, "n": 2}

Only one event carries target-policy mass, despite two observed events.

def policy_report(events, max_weight=None):
    raise NotImplementedError