##What artificial agents can teach us about social decisions
John Pearson
Biostatistics and Bioinformatics
[pearsonlab.github.io/artificial-agents-social-decisions](https://pearsonlab.github.io/artificial-agents-social-decisions)
### Pros:
- Well-studied, normative solutions
### Cons:
- Highly idealized, limited dynamics
- Biologically aligned?
- Doesn't matter if it's social
- Requires anticipating another agent
- Repeatable, but lots of variation
- Decisions tightly linked to movement
Kelsey McDonald
Scott Huettel
Penalty Shot
Your browser does not support the video tag.
McDonald, Broderick, Huettel, Pearson
- two monkeys, shooter and goalie (shooter recorded)
- controlled by joysticks
- roles rotated, animals know each other
- repeated sessions
- rapidly learned, rich dynamics
Your browser does not support the video tag.
How it works
Human, computer opponents
Incentive compatible
Coevolving strategies
Highly variable strategies
How do we model this?
$$p(\text{change}) = \Phi(f(s(t), \omega(t)))$$
$$f(s(t), \omega(t)) \sim \mathrm{GP}(0, k)$$
Informally:
strategy depends on game state $s$, opponent $\omega$
$\mathrm{GP}$ is a distribution over functions
Bayesian model: we can quantify uncertainty
Joint model: can do "counterfactual" experiments
How likely are you to win?
Your browser does not support the video tag.
How likely are you to change course?
Your browser does not support the video tag.
Real and predicted change points
How opponent-sensitive are you?
Your browser does not support the video tag.
Does identity matter?
Of course!
But context also matters.
Can we separate identity from context?
Major challenge in social neuro/pscyh
Does identity matter?
Summary
Flexible models let us study more realistic decisions
Efficient models let us make individualized predictions
Accurate models (artificial agents) let us ask "what if" questions, perform synthetic experiments