Learning to GigWorkforce Development under Worker Agency

We study a novel sequential decision-making problem inspired by gig economy platforms, where workers autonomously choose whether to accept tasks and improve with experience. Unlike traditional recommendation or bandit settings, platforms in this context cannot directly assign jobs; they must recommend a subset of workers and rely on their voluntary response. We model this as a two-phase bandit problem in which arms (workers) exhibit agency, stochastic task acceptance, and endogenous learning, where performance improves with experience. These features introduce a new exploration-exploitation trade-off: choosing experienced workers for short-term reward versus investing in newer workers to enhance future performance. We show that computing the optimal policy is intractable due to the highdimensional state space. We propose a greedy benchmark that captures immediate expected utility, and design a learning algorithm, C-UCB, that achieves sublinear O(log T) regret with respect to this benchmark. Our framework highlights key algorithmic challenges in human-in-the-loop assignment systems and offers foundations for recommendation systems that support both autonomy and skill development.

From SSRN, posted as "Two-Phase Bandits with Arm Agency and Learning"