Goal-directed decision making in prefrontal cortex: A computational framework
Matthew Botvinick and James An, Princeton University

Poster M86

Behavioral and neuroscientific research suggests that action selection occurs within two distinct systems. In one, actions are selected based on pre-established situation-action associations. In the other, they're selected in a prospective, goal-directed manner. Goaldirected action selection is believed to depend on both dorsolateral and orbital prefrontal cortex, but the computations involved are not well understood.

Dorsolateral PFC Motor cortices Parietal cortex Medial temporal cortex Orbitofrontal cortex

 a s

m

 a s u s

m

 a s u

m

 a s u

m

s u
uG

1.0

Left (S2)

Left (S1)

0.5

EV

0.75

Right (S2)
0.0 1

Right (S1)
0.64 50

Iteration

s6 s4
B

1.00

Left at s2

p()

s1

s2
A

s3

s5

EV

0.5277

0.33

s7

s8

s9
0.00 1

Straight at s2

Right at s2

Iteration

0.5130 400

We consider an account of goal-directed decision making that models it as a form of probabilistic inference. Policies, actions, and the transition and reward functions are represented within a probabilistic graphical model, whose components map onto specific neuroanatomic structures. Action selection is performed by querying this model, using Bayesian inference to reason from rewards to policies. We introduce a recursive inference procedure that provably converges on optimal policies, and apply this in simulations of some key findings from animal and human research.

p(uG)

p()

p(uG)