A Synthetic Comparison of Local and Centralized-Critic PPO for Automated Job Search
Aim: This study provides a framework for automated job search and compares the performance of parameter-shared Proximal Policy Optimization (PPO) methods using local and centralized critics with random, skill-matching-based greedy hiring, and congestion-aware hiring policies. It extends previous research on PPO-based hiring through additional evaluation conditions, methods, and in-depth performance assessment. Methods: The study employed a simulated 12-round, partially observable hiring environment involving 16 applicants and stochastic hiring. Nine fixed conditions were established based on the number of initial vacancies and the congestion coefficient. Following validation-based hyperparameter tuning, each method was trained using 30 random seeds and evaluated on 256 different seeds for each condition. Results: The simulator-level hire rates were 80.64% for greedy hiring, 80.79% for LocalPPO, and 80.77% for CentralPPO. Both PPO methods outperformed the random and congestion-aware heuristic approaches in terms of mean reward. Compared with the greedy baseline, LocalPPO and CentralPPO improved the hire rate by 0.155 and 0.134 percentage points, respectively, although both recorded slightly lower mean rewards. The results also revealed minimal performance differences between local and centralized critics. Conclusion: The findings suggest that PPO-based approaches can offer competitive performance compared with conventional hiring strategies. However, the simulation results cannot be assumed to generalize to real-world hiring environments without further validation. Recommendation: Future research should investigate different policy classes and market configurations using hold-out data. Further validation in real-world hiring settings is also necessary to establish the applicability and generalizability of the proposed framework.
Authors
- Sai Pavan Kumar Gopularam
Publication Details
- Journal
- American Journal of Technology
- Published
- 2026-11-21
- DOI
- https://doi.org/10.58425/ajt.v5i13.649
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- article
- Field-Weighted Citation Impact
- 0.00