Multi-Objective Bandit Algorithm for Pareto Frontier Approximation
A new paper on arXiv introduces THV-UCB, an algorithm for multi-objective slate selection in stochastic bandit problems. At each round, the agent selects k arms and observes d-dimensional reward vectors under semi-bandit feedback. Instead of identifying a single optimal arm, the goal is to maintain a small set of actions approximating the Pareto frontier. The objective is formalized through dominated hypervolume induced by the selected subset, with an α-approximate hypervolume regret where α = 1 - 1/e reflects the greedy maximization guarantee for monotone submodular functions. THV-UCB selects arms greedily based on optimistic estimates of marginal hypervolume contributions. The algorithm achieves a gap-free regret bound of Õ(d√(nkT)).
Key facts
- arXiv paper 2607.26273 introduces THV-UCB algorithm
- Addresses stochastic multi-objective bandit problem with slate selection
- Agent selects k arms per round, observes d-dimensional reward vectors
- Objective is to approximate Pareto frontier with a small set of actions
- Uses dominated hypervolume as performance metric
- α-approximate hypervolume regret with α = 1 - 1/e
- THV-UCB selects arms greedily based on optimistic marginal hypervolume contributions
- Regret bound: Õ(d√(nkT))
Entities
Institutions
- arXiv