Decision Tree Tool dtControl2+ε Trades Optimality for Explainability in MDPs
A new extension of the dtControl2 tool, called dtControl2+ε, allows for smaller and more interpretable decision trees representing controllers in Markov decision processes by accepting a bounded imprecision ε. The tool guarantees ε-optimality while reducing tree size by orders of magnitude, making complex controllers more human-comprehensible. This addresses the challenge of large systems with many corner cases where full decision trees become too complex. The work is published on arXiv under identifier 2607.25925.
Key facts
- dtControl2+ε extends dtControl2 with ε functionality
- Accepts imprecision ε ≥ 0 to reduce decision tree size
- Guarantees ε-optimality of the controller
- Decision trees are orders of magnitude smaller
- Aims to improve explainability of controllers in MDPs
- Published on arXiv with ID 2607.25925
- Addresses complexity in large systems with many corner cases
Entities
Institutions
- arXiv