Exploration and Generalization Over the Lifespan

• In Event: The Development of Risk Preferences and Exploration Strategies in Human Children and Chimpanzees

Abstract

Learning is not static over the lifespan, but may take different forms in childhood, adolescence, and adulthood (Moran et al., 2014). This could be due to differences in cognitive development but also due to unique opportunities at different stages of life (Lindenberger & Lövdén, 2019; Nussenbaum & Hartley, 2019). Childhood offers a rich environment for learning new skills, while adolescence often provides individuals with their first experiences of individual freedom. How does learning and exploration change over the lifespan? What changes can we attribute to differences in cognitive development and what represents a lifespan rational adaptation to age-related changes in our learning environment?
Combining data across 5 different experiments and from 602 participants between the ages of 4 and 64, we plot a detailed trajectory of how exploration and generalization evolves over the lifespan. All experiments used variations of a spatially correlated bandit task (Wu et al., 2018), where participants acquired rewards by repeatedly choosing from at least 64 different options. Each option provided a normally distributed reward, where the expected reward was spatially correlated such that nearby options had similar rewards. This spatial structure allowed for statistical inference and generalization, while the rich decision space allowed us to differentiate between uncertainty directed exploration and undirected, random exploration.
Using a combination of behavioral analyses and computational models, we provide insights into age-related differences in value-directed learning and exploration. While we find support for existing hypotheses suggesting young children are more prone to random exploration (Gopnik et al., 2017), this was not the full picture. Already in children as young as 4, we see the emergence of uncertainty directed exploration, suggesting that more sophisticated and goal-directed forms of exploration come online earlier than was previously believed. In comparison, adults had less directed exploration than children, but equal levels of random exploration. And although adults generalized more than children, they nevertheless showed an intriguing bias towards undergeneralization, corresponding to an underestimation of the level of spatial correlations in the reward structure. However, simulations suggest that this is often a beneficial bias relative to overgeneralization, and sometimes even outperforms an exact match to the ground truth. The adolescent sample has not yet been analyzed, but we predict they will have the highest levels of directed exploration (van den Bos et al., 2012), but with the same amount of random exploration (Somerville et al., 2017).
Altogether, our results offer unique insights into the dynamics of learning and exploration over the lifespan. Our use of computational models allow us to compare participant data against counter-factual simulations using alternative parameterizations, in order to determine what is rational and what is still developing. Ultimately, this work seeks to contribute to growing literature on how curiosity and intrinsic motivation both shapes and adapts to our informational landscape.

Authors

• Charley M Wu, University of Tübingen
  Presenting Author

• Anna Giron, University of Tübingen
  Non-Presenting Author

• Eric Schulz, Max Planck Institute for Biological Cybernetics
  Non-Presenting Author

• Björn Meder, Health and Medical University Potsdam
  Non-Presenting Author

• Azzurra Ruggeri, Max-Planck Institute for Human Development
  Non-Presenting Author

• Simon Ciranka, Max Planck Institute for Human Development
  Non-Presenting Author

• Wouter Van den Bos, University of Amsterdam
  Non-Presenting Author