Adaptive Smooth Tchebycheff Attention for Multi-Objective Policy Optimization
A new framework called Adaptive Smooth Tchebycheff (AST) addresses the challenge of balancing conflicting objectives in multi-objective reinforcement learning for robotics. Linear scalarization methods offer stability but cannot recover solutions in non-convex regions of the Pareto front. Static non-linear scalarizations like Tchebycheff can access these regions but suffer from gradient variance and instability in deep RL. AST dynamically modulates optimization landscape curvature using a conflict-driven controller that adjusts smoothness based on real-time gradient interference. This enables precise non-convex scalarization when objectives align and reverts to stable approximations when gradients conflict. The approach is detailed in arXiv:2605.12771.
Key facts
- arXiv:2605.12771
- Announce Type: cross
- Adaptive Smooth Tchebycheff framework proposed
- Conflict-driven controller regulates optimization smoothness
- Addresses non-convex trade-offs in multi-objective RL
- Linear scalarization cannot recover non-convex Pareto front solutions
- Static Tchebycheff suffers gradient variance in deep RL
- Method anneals toward precise scalarization when objectives align
Entities
Institutions
- arXiv