In the chess example, the player might have learned from past experience that sacrificing one's queen is generally a bad choice. They would then be less likely to consider solutions involving this move, and less likely to make the 'aha' discovery that it is indeed optimal to do so in this case.

This strategy enables efficient decision-making, but can also lead us astray when our past experience isn't informative for the current context.

We suggest that humans and AI often fail on these problems because they both rely on a planning architecture where possible options are generated based on what was successful in the past and these options are then evaluated based on knowledge of the task at hand.

An example of such a problem is a chess puzzle where sacrificing one's queen is necessary to ensure victory a few steps down the line. A novice chess player who understands the rules of chess might never consider this solution, though they would certainly select it if it were proposed to them

TLDR: kids are smart

Our results provide an account of how children and adolescents make flexible choices in a changing world, and suggest a need to better understand how diverse learning strategies influence choices across development.

Here, we ask if children can rely on alternative strategies to use structured knowledge in their choices. We show that children use simplified predictive representations like the Successor Representation to efficiently predict the likely outcomes of their choices without multi-step simulation.

One possible reason for this is that using structured knowledge to make decisions often involves mentally simulating multi-step actions and their outcomes, which can be computationally costly and depend on still-developing cognitive processes