Reinforcement Learning for Scalable and Trustworthy Intelligent Systems

A dissertation presents research on scaling reinforcement learning across distributed systems while ensuring trustworthy behavior in AI applications. The work addresses communication efficiency in federated settings and alignment with human preferences in large language models, proposing that next-generation intelligent systems require both optimization efficiency and safety mechanisms.

This research tackles two interconnected challenges limiting practical deployment of reinforcement learning at scale. As AI systems become increasingly distributed and powerful—particularly in post-training large language models and autonomous agents—the technical community faces pressure to solve scalability and safety simultaneously rather than sequentially.

The federated learning approach addresses real infrastructure constraints: communication bandwidth limitations and heterogeneous computational resources across distributed agents create bottlenecks in current systems. Traditional centralized approaches become impractical when training spans multiple organizations or edge devices. The dissertation's focus on asynchronous federated optimization directly responds to these deployment realities, suggesting the field is moving beyond theoretical frameworks toward production-ready systems.

The trustworthiness component reflects growing concern about AI safety in high-stakes applications. Privacy-aware information disclosure and human preference alignment represent critical requirements for deploying autonomous systems in regulated environments. As large language models become infrastructure for enterprise and consumer applications, unaligned behavior poses both reputational and liability risks.

For the AI industry, this research validates that scalability and safety are complementary rather than competing objectives—solving one without the other produces systems unsuitable for deployment. The federated optimization contributions could accelerate multi-organizational AI development, while the safety framework helps establish governance patterns for autonomous systems. The work also signals academic consensus that reinforcement learning remains fundamental to advanced AI development, contrasting with narratives suggesting alternatives.

• →Federated reinforcement learning optimization addresses communication and computational constraints in distributed AI systems
• →Human preference alignment and safety mechanisms are becoming prerequisites rather than optional features in modern AI deployment
• →Asynchronous optimization enables practical training across heterogeneous agents with limited bandwidth
• →Privacy-aware information disclosure in language models requires contextual safety constraints during reinforcement learning
• →Scalability and trustworthiness must be co-designed to create production-ready intelligent systems