| Definition | A CIRL Benchmark with Variational Inverse Algorithm for Constrained Reinforcement Learning.” |
| No of environments | 1 |
An introduction of the Constrained Inverse Reinforcement Learning (CIRL), featuring modified MuJoCo environments and a self-driving car-inspired environment. To address challenges in this domain, a novel algorithm is proposed: Variational Inverse Constrained Reinforcement Learning (VICRL), a Bayesian-based approach designed to solve CIRL problems by incorporating uncertainty into the learning process. The benchmark itself provides valuable insights into the problem space and highlights areas for development.
The proposed Bayesian approach is a novel contribution, and the work fills an important gap by offering a precise goal.