Accelerated Adaptive Decision Making for Autonomous Agents: Optimization and Coverage

Westudyadaptivedecisionmakingunderuncertaintyforautonomous agents that interact with expensive black-box evaluators under strict evaluation budgets. These settings arise naturally in agentic scientific discovery where an agent must sequentially choose what to test next while balancing exploration and exploitation and respecting safety or feasiblity constraints in the objective space. A key theme of my dissertation is acceleration: characterizing how faster estimation primitives can fundamentally improve sample efficiency, and designing algorithms that exploit such acceleration in principled ways. As initial results, (i) for sequential decision making in non-linear bandits, we propose Q-NLB-UCB, a new input dimension-free quantum non-linear bandit algorithm with 𝑂(polylog𝑇) regret, (ii) for adaptive constrained objective-space coverage, we further formulate the new multi-objective coverage (MOC) problem where our goal is to identify a small set of representative samples whose predicted outcomes broadly cover the feasible multi-objective space. To solve this MOC problem, we propose a novel search algorithm, MOC-CAS, which employs an upper confidence bound-based acquisition function to select optimistic samples guided by Gaussian process posterior predictions. Empirically, MOC-CAS achieves superior performances over competitive baselines on large-scale protein-target datasets. Going forward, we will develop acceleration-aware theory for optimistic exploration under explicit cost models, study multi-agent extensions for parallel experimentation, and extend to graph-structured action spaces using uncertainty-aware learned representations.

Authors

• Zakaria Shams Siam University at Albany, State University of New York

Keywords

• Adaptivedecisionmaking
• Banditoptimization
• Constrainedobjectivespace coverage
• Autonomous scientific discovery