Boltzmann-Softmax Control for Fair AI Resource Allocation
A novel framework named Computable Fair Division (CFD) offers a new perspective on the Boltzmann-Softmax function, viewing it as a probabilistic mechanism for resource distribution in extensive AI systems. The inverse temperature parameter, β, is redefined to serve as a computable control variable that regulates the balance between efficiency and fairness. Through static analysis, a Pareto frontier is identified, featuring a nearly optimal Stability Corridor where total loss stays roughly constant across various policy weights. In dynamic environments, the Adaptive Hard-Cap Controller++ (AHC++) adjusts β in real-time, utilizing the discrepancy between observed dominance and a target set by the policy for feedback, thereby tackling dominance concentration that threatens system diversity and stability.
Key facts
- CFD reinterprets Boltzmann-Softmax as probabilistic resource allocation
- β is redefined as a computable control variable for efficiency-fairness balance
- Static analysis shows a Pareto frontier with Stability Corridor
- AHC++ updates β in real time using dominance error feedback
- Conventional policies focus on efficiency metrics leading to dominance concentration
- The framework targets GPU compute time and bandwidth allocation
- Simulations demonstrate the approach
- The paper is from arXiv:2605.22827
Entities
Institutions
- arXiv