m‾(st−1) which were exponential moving averages of choices following trials ending in state st−1, updated as:14m‾(st−1)←(1−αm)m‾(st−1)+αm(c−0.5)Net action values were given by a weighted sum of model-free, motor-level model-free and model-based action values, biases and perseveration.15Qnet(c)=GmfQmf(c)+GmoQmo(c,st−1)+GmbQmb(c)+X(c)Where Gmf, Gmo and Gmb are weights controlling the influence of respectively the model-free, motor-level model-free and model-based action values, and X(c) is biases and perseveration where:16X(top)=Bc+Br(st−1−0.5)+Pcc‾+Pmm‾17X(bottom)=0where st−1 = 1 if previous second step state is left and 0 if right.
