Past studies on exploration predominantly used multi-armed-bandit paradigms and mostly focused on simple concepts such as uncertainty-guided exploration. In our work, we argue that by using games to study exploration we can observe and investigate more elaborated strategies like fun and empowerment.

Tool use is a defining aspect of human intelligence, but traditional studies of tool use require time-consuming, in person experiments. Here we present an alternative paradigm - the Virtual Tools Game - which taps into aspects of human cognition that support tool use and physical reasoning more generally, and allows for more detailed, quantitative analyses of behavior. I will show how we’ve used this domain to study and build computational models of how people solve new tool use puzzles rapidly, and how they learn relational tool using strategies, or affordances, from very limited amounts of experience.

Tetris is one of the most well known and most liked games of all time. In this talk, we will examine the sequence of decisions that are encountered in playing the game. We will show that these decision tasks display some characteristics that make decision making “easy” in a well-defined sense. In conclusion, we will discuss what the results might reveal about human and artificial intelligence.

Empirically, only a negligible fraction of humanity can be enticed into research participation through cash or course credit. Among that fraction, motivation is often low, as is data quality. We discuss the research advantages of designing intrinsically motivating experiments.

Cognitive scientists have long known that people’s mental representations of games and problems shape how they solve them. But, what shapes these representations? In our talk, we will discuss several recent projects that indicate people form resource rational planning representations when they play games. These findings provide a point of departure for revisiting classic questions about the interplay of representation and computation in human problem solving.

Humans are capable of learning a wide array of complex decision making tasks with remarkable sample efficiency. For example, when learning to play a new video game, humans often make use of high level assumptions about the structure of games in general (e.g., most games have at least one way to win and one way to lose, objects in a game tend to serve a purpose, etc.) to more precisely direct exploration. However, the precise nature of these high level assumptions remains unknown. In this work we present a computational account of how high level assumptions about video game structure can act as inductive biases to guide search through theory-space and ultimately facilitate model-based exploration.

Experimental reductionism has served cognitive science well, but could limit understanding of cognitive processes with real-world complexity. Games could help to overcome these limitations. We used a variant of tic-tac-toe that has ~10^16 states, yet allowed us to fit a computational model, inspired by best-first search, to human moves. This model predicts moves in new positions, decisions in new tasks, eye fixation patterns, mouse movements, and response times. Using the model, we characterized the effects of time pressure and expertise. Finally, we collected ~159 million moves of the same game on a mobile platform, providing insights into complex planning in the wild.

Drawing is one of the most basic tools humans have for encoding our thoughts in a durable format, enlarging our collective capacity to imagine, think, and solve problems. It's also a heck ton of fun. In this talk, I will describe why I think games that elicit open-ended behaviors are so valuable in cognitive science, and some of what we have learned by embracing drawing games in my lab's research.

Through a process of cultural accumulation, humans have developed vast technological repertoires that have enabled us to survive everywhere from the tundra to the Earth's orbit. How did these technological repertoires come to be? In this talk, I will explore this question by analyzing behavior in One Hour One Life, a multiplayer online game where players can build technologically complex communities over many generations (N = 25,060 players, 30,755 communities, 500,693 lives lived).

Is it possible to compare children with artificial agents in unified environments? Is it possible to adapt an environment designed to test the behavior of artificial agents eg. the DeepMind lab environment, for use with real children? We have developed a platform and framework based on DeepMind Lab - which is a first person 3D navigation and puzzle-solving environment originally built for testing agents in mazes with rich visuals to be used as a unified environment with both children and agents. We also look at adapting a classic developmental Blicket Detector task to be tested virtually with children and agents and studying these games casually.