Skip to content

Implementations and examples of common offline policy evaluation methods in Python.

License

Notifications You must be signed in to change notification settings

banditml/offline-policy-evaluation

Repository files navigation

Offline policy evaluation

PyPI version Downloads

Implementations and examples of common offline policy evaluation methods in Python. For more information on offline policy evaluation see this tutorial.

Installation

pip install offline-evaluation

Usage

from ope.methods import doubly_robust

Get some historical logs generated by a previous policy:

df = pd.DataFrame([
	{"context": {"p_fraud": 0.08}, "action": "blocked", "action_prob": 0.90, "reward": 0},
	{"context": {"p_fraud": 0.03}, "action": "allowed", "action_prob": 0.90, "reward": 20},
	{"context": {"p_fraud": 0.02}, "action": "allowed", "action_prob": 0.90, "reward": 10},
	{"context": {"p_fraud": 0.01}, "action": "allowed", "action_prob": 0.90, "reward": 20},     
	{"context": {"p_fraud": 0.09}, "action": "allowed", "action_prob": 0.10, "reward": -20},
	{"context": {"p_fraud": 0.40}, "action": "allowed", "action_prob": 0.10, "reward": -10},
 ])

Define a function that computes P(action | context) under the new policy:

def action_probabilities(context):
    epsilon = 0.10
    if context["p_fraud"] > 0.10:
        return {"allowed": epsilon, "blocked": 1 - epsilon}    
    return {"allowed": 1 - epsilon, "blocked": epsilon}

Conduct the evaluation:

doubly_robust.evaluate(df, action_probabilities)
> {'expected_reward_logging_policy': 3.33, 'expected_reward_new_policy': -28.47}

This means the new policy is significantly worse than the logging policy. Instead of A/B testing this new policy online, it would be better to test some other policies offline first.

See examples for more detailed tutorials.

Supported methods

  • Inverse propensity scoring
  • Direct method
  • Doubly robust (paper)