Research map / Reinforcement learning / Policy gradient and actor-critic
Integrating Reward Maximization and Population Estimation: Sequential Decision-Making for Internal Revenue Service Audit Selection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ABS with Horvitz–Thompson population estimation instead of greedy selection with model-based estimation | IRS audit selection · IRS NRP | population estimate absolute mean percent difference: 16.4 -> 0.4 | Helped |
| Use ABS-1 instead of ε-only random selection for an unbiased population estimate | IRS audit selection · IRS NRP | cumulative reward: $41.3M* -> $41.5M* | No clear effect |
| Use ABS rather than the best ε-only unbiased estimator | IRS audit selection · IRS NRP | population estimate percent difference standard deviation: 16% lower variance | Helped |
| Add even a 10% random sample to greedy selection | IRS audit selection · IRS NRP | population estimate bias: more than 2.5× reduction | Helped |
| Use model-based population estimation without randomized sampling | Greedy, ε-greedy, and UCB · IRS NRP | population estimate bias: Greedy: 16.4; UCB-1: 15.3; UCB-2: 15.6 | Hurt |
| Increase the ABS reward emphasis using more reward-optimal hyperparameter settings (ablation) | ABS · IRS NRP | population estimate variance | Hurt |
| Add trimming to ABS to impose a minimum sampling probability across bins (ablation) | ABS · IRS NRP | cumulative reward: 3.2x reward | Helped |
| Use Random Forest Regression rather than LDA for reward prediction | IRS audit selection · IRS NRP | cumulative reward: RF Greedy selects between $1.8M and $9.4M more cumulative TPI than LDA Greedy | Helped |
| Use optimism-based UCB rather than RF Greedy | IRS audit selection · IRS NRP | cumulative reward: $42.4M* vs $43.6M* | No clear effect |
| Use UCB rather than ABS-1 for the income prediction task | CPS dataset · CPS | Reward: 473 vs 431 | Helped |
| Use ABS-1 rather than UCB for the income prediction task | CPS dataset · CPS | Bias: 13.4 -> 0.2 | Helped |
| Use Thompson sampling rather than ABS-1 for the income prediction task | CPS dataset · CPS | Reward: 427.2 vs 431 | No clear effect |
Closest papers by meaning
- RR-PU: A Synergistic Two-Stage Positive and Unlabeled Learning Framework for Robust Tax Evasion Detection (AAAI 2024 · similarity 0.82)
- On Mitigating Affinity Bias through Bandits with Evolving Biased Feedback (ICML 2025 · similarity 0.82)
- A General Framework for Off-Policy Learning with Partially-Observed Reward (ICLR 2025 · similarity 0.81)
- From Past to Future: Rethinking Eligibility Traces (AAAI 2024 · similarity 0.81)
- Log-Sum-Exponential Estimator for Off-Policy Evaluation and Learning (ICML 2025 · similarity 0.81)
- Policy Learning for Balancing Short-Term and Long-Term Rewards (ICML 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.