Table 1.
Model fits.
| Scale factors | |||||||
|---|---|---|---|---|---|---|---|
| Model | Model fit (LLE) | Corrected model fit (AIC) | Offset (C0) | Signed prediction error (Cδ) | Unsigned prediction error (C|δ|) | Expected value (CV) | Learning rate (α) |
| 20 MIN DELAY | |||||||
| Baseline (no reward influence) | 70.755 ± 1.928 | 143.510 ± 3.856 | 0.721 ± 0.075 | – | – | – | – |
| Signed prediction error (δ) | 60.440 ± 1.636 | 126.881 ± 3.272 | 0.735 ± 0.071 | 0.369 ± 0.158 | – | – | 0.395 ± 0.077 |
| Unsigned prediction error (|δ|) | 61.718 ± 1.989 | 129.436 ± 3.979 | 0.562 ± 0.107 | – | 0.503 ± 0.166 | – | 0.583 ± 0.074 |
| Expected value (V) | 61.593 ± 2.124 | 129.186 ± 4.247 | 0.584 ± 0.104 | – | – | 0.501 ± 0.221 | 0.372 ± 0.064 |
| Signed prediction error and expected value (δ+V)* | 57.828 ± 1.978 | 123.656 ± 3.957 | 0.350 ± 0.132 | 0.725 ± 0.154 | – | 1.156 ± 0.180 | 0.329 ± 0.057 |
| Unsigned prediction error and expected value (|δ|+V) | 59.625 ± 2.143 | 127.250 ± 4.287 | 0.530 ± 0.099 | – | 0.361 ± 0.212 | 0.360 ± 0.239 | 0.456 ± 0.075 |
| 24 H DELAY | |||||||
| Baseline (no reward influence) | 76.445 ± 1.723 | 154.889 ± 3.447 | 0.496± 0.026 | – | – | – | – |
| Signed prediction error (δ) | 64.079 ± 1.625 | 134.159 ± 3.250 | 0.507 ± 0.073 | 0.215 ± 0.171 | – | 0.333 ± 0.068 | |
| Unsigned prediction error (|δ|) | 65.270 ± 2.201 | 136.539 ± 4.401 | 0.417 ± 0.108 | – | 0.292 ± 0.202 | – | 0.404 ± 0.085 |
| Expected value (V) | 64.267 ± 1.968 | 134.535 ± 3.937 | 0.402 ± 0.117 | – | – | 0.268 ± 0.265 | 0.456 ± 0.080 |
| Signed prediction error and expected value (δ+V)* | 61.165 ± 1.985 | 130.329 ± 3.970 | 0.223 ± 0.123 | 0.469 ± 0.173 | – | 0.729 ± 0.274 | 0.312 ± 0.066 |
| Unsigned prediction error and expected value (|δ|+V) | 62.943 ± 2.291 | 133.887 ± 4.582 | 0.284 ± 0.146 | – | 0.327 ± 0.199 | 0.277 ± 0.257 | 0.433 ± 0.078 |
LLE is the negative log-likelihood estimate. AIC is Akaike's Information Criterion. The scale factors Cδ, C|δ|, and CV determine the contribution of signed prediction errors, unsigned prediction errors, and expected value to memory performance, respectively. α is the learning rate for the Q-learning rule.
*
Denotes the model providing the best fit to behavioral data, as indicated by significantly lower AIC values (see Section Computational Approach). Mean ± SEM.