In order to find extrinsic rewards, humans explore their environment even if exploration requires several intermediate, reward-free decisions. It has been hypothesized that intrinsic rewards, such as novelty, surprise, or information gain, guide this reward-free exploration. However, in artificial agents, different intrinsic reward signals induce exploration strategies that respond differently to stochasticity. In particular, some strategies are vulnerable to the “noisy TV problem,” i.e., an attraction to irrelevant stochastic stimuli. Here, we ask whether humans exhibit a similar attraction to reward-free stochasticity. We design a multistep decision-making paradigm in which participants search for rewarding states in a complex environment containing a highly stochastic but reward-free subregion. We show that i) participants persistently explore the stochastic subregion, and ii) their decisions are best explained by a novelty-driven exploration strategy, compared to alternatives driven by information gain or surprise. Our findings suggest that novelty and extrinsic rewards jointly control human exploration in complex environments.
Modirshanechi et al. (Thu,) studied this question.