model-free learning
Decision-making that relies on trained habits—attaching value directly to actions based on past reward—without using a model of how actions lead to outcomes; corresponds to Kahneman's system one.
As defined in 176 | Joshua Greene on Morality, Psychology, and Trolley Problems
Where it comes up · 1
Most important claims · 3 of 3Search all claims for “model-free learning” →
- In Cushman and Patel's two-stage navigation game—which has nothing to do with morality—people who rely more on model-based reasoning (using a map) rather than trained habits (model-free) are more likely to show a utilitarian pattern in their moral judgments, demonstrating the same neurobiology operates in maze navigation in humans and rats and in moral judgment.176 | Joshua Greene on Morality, Psychology, and Trolley Problems
- The negative emotional reaction to pushing someone off a footbridge comes from basic reinforcement learning—the same model-free mechanism seen in deep reinforcement learning neural networks and in rats, monkeys, and humans—whereby we learn in toddlerhood that actions like pushing people are not okay, so negative reactions become attached to them.176 | Joshua Greene on Morality, Psychology, and Trolley Problems
- The dual-process model of moral reasoning is the moral analogue of Kahneman's Thinking Fast and Slow: model-free learning corresponds to fast system one (attaching value directly to actions) and model-based learning to slow system two (attaching value to actions via understanding causal connections to consequences), and this applies to mental actions too, since reasoning follows reinforced habits of thought.176 | Joshua Greene on Morality, Psychology, and Trolley Problems