Machine Learning Optimization Methods

This paper introduces Bounded Ratio Reinforcement Learning (BRRL), a theoretical framework that bridges the gap between trust region methods and PPO's clipped objective, leading to a new algorithm called Bounded Policy Optimization (BPO) that provides monotonic improvement guarantees while matching or exceeding PPO's performance. The framework also extends to Group-relative BPO (GBPO) for large language model fine-tuning.

GSQ is a new scalar quantization method for large language models that uses Gumbel-Softmax relaxation to jointly optimize grid assignments and scales, achieving accuracy comparable to complex vector quantization methods while remaining compatible with existing inference kernels. It successfully quantizes models to 2-3 bits per parameter and scales to trillion-parameter mixture-of-experts models.

This note demonstrates that TurboQuant, a recent quantization method, is a suboptimal special case of the earlier EDEN/DRIVE quantization schemes. EDEN consistently outperforms TurboQuant across all tested scenarios, often by more than one bit of precision.

This paper introduces a bandit algorithm for learning smooth functions over graphs, where neighboring nodes have similar payoffs. The approach achieves regret bounds that scale with an effective dimension rather than the total number of nodes, enabling efficient online learning for applications like content recommendation with thousands of items from just tens of evaluations.

This paper develops a graph neural network approach to classify bridges based on their disaster-resilience roles by analyzing metapaths connecting highways, bridges, and critical buildings. The method helps prioritize bridge maintenance budgets by identifying which bridges are essential for supply chains, medical access, or residential protection during disasters.

BACO is a framework that compresses embedding tables in recommender systems by clustering similar users and items to share embeddings, achieving over 75% parameter reduction with minimal accuracy loss. It outperforms existing methods by being up to 346X faster while maintaining recommendation quality.

This study develops a machine learning model using molecular descriptors to screen over 7,000 natural compounds for potential anti-Alzheimer activity, identifying 73 promising candidates. The cheminformatics approach demonstrates moderate predictive performance and highlights key molecular features associated with therapeutic potential.

Platypoos is a planning algorithm for deterministic environments with stochastic rewards that automatically adapts to unknown reward scales and smoothness without requiring prior knowledge of discount factors or reward bounds. It achieves optimal sample complexity with matching upper and lower bounds.

This paper introduces the BlockEncoding interface in the Eclipse Qrisp framework, which makes block-encoding techniques accessible as high-level programming abstractions for implementing advanced quantum algorithms. The interface simplifies the practical implementation and resource estimation of algorithms like QSVT, QSP, and Hamiltonian simulation.