Example Codes

SCOPE-RL

See also

Please also refer to SCOPE-RL package reference for APIs.

Basic and High-Confidence Off-Policy Evaluation (OPE):

Basic OPE

• Logged datasets and inputs

• Basic Off-Policy Evaluation (DM, PDIS, DR, ..)

• Marginal Off-Policy Evaluation (SMIS, SMDR, SAMIS, SAMDR, ..)

• High-Confidence Off-Policy Evaluation (Hoeffding, Bernstein, ..)

• Extention to continuous action space

Cumulative Distribution OPE (CD-OPE):

Cumulative Distribution OPE

• Logged datasets and inputs

• Estimating Cumulative Distribution Function

• Estimating risk-functions (mean, variance, CVaR, ..)

Off-Policy Selection

Off-Policy Selection (OPS)

• OPS via Basic OPE

• OPS via Cumulative Distribution OPE

• Obtaining oracle selection results

Assessing OPE Estimators

Off-Policy Selection (OPS)

• Conventional “accuracy” metrics

• Top-\(k\) risk-return tradeoff metrics

• Validation visualization

Implementing Custom OPE Estimators:

Basic OPE (Continuous)

• Custom Basic OPE estimators

• Custom Cumulative Distribution OPE estimators

Handling Multiple Datasets:

Basic OPE (Continuous)

• Logged datasets and inputs

• (Basic) Off-Policy Evaluation

• Cumulative Distribution Off-Policy Evaluation

• Off-Policy Selection

• Assessments of OPE and OPS

Handling Real-World Datasets:

Basic OPE (Continuous)

• Logged dataset

• Input dict

See also

For the data collection and integration with d3rlpy in policy learning, please also refer to this page.

See also

The comprehensive quickstart examples with the provided sub-packages are available in the GitHub repository:

• Quickstart with RTBGym

• Quickstart with RECGym

• Quickstart with BasicGym

<<< Prev Quickstart

Next >>> Basic OPE