Oui, Chef!!
A restaurant management game in which machine learning is the central mechanic, not a background implementation detail. The player's job is to train the kitchen staff — neural network agents — to correctly prepare dishes by demonstrating the mapping from recipe to ingredients. The cooks learn through supervised training; the player's interactions are the training data.
Built with Quentin Petraroia and Sam Lee in our third year of undergraduate Computing at Queen's, and presented at Graphics Interface 2019.
Two design decisions turned out to be more interesting than expected. First, the training interface was simple enough that players with no machine learning background could intuitively understand what they were doing — accept a cook's attempt, reject it, repeat. The cognitive model players built of the training process matched the actual process well enough to produce effective training behaviour. Second, catastrophic forgetting — the tendency of neural networks to overwrite earlier learning when trained on new examples — made the cooks less predictable and, counterintuitively, more believable. A cook who sometimes forgot a recipe they had learned behaved more like a person than a cook with perfect recall would have.
This was the first project in which I thought carefully about how a player's mental model of an AI system shapes their interaction with it — and about how apparent failures of the AI can be features rather than bugs depending on what the player understands them to mean.
Read the paper →