Predicting human decision-making across task conditions via individuality transfer
- The University of Osaka, Japan
Figures
Figure 1
The EIDT (encoder, individual latent representation, decoder, and task solver) framework for individuality transfer across task conditions.
The encoder maps action(s) α, provided by an individual performing a specific problem in the source task condition , into an individual latent representation (represented as a point in the two-dimensional space in the center). The individual latent representation is then fed into the decoder, which generates the weights for a task solver. The task solver predicts the behavior of the same individual in the target task condition . During the training, a loss function evaluates the discrepancy between the predicted behavior and the actual recorded behavior β of individual . The encoder’s input is referred to as an action sequence, the form of which depends on task. For example, in a sequential Markov decision process (MDP) task, an action sequence consists of an environment (state transition probabilities) and a sequence of actions over multiple episodes. For a digit recognition task, it consists of a stimulus digit image and the corresponding chosen response.
Figure 2
Comparison of prediction performance in Within-Condition Prediction for the MDP task.
The plots show the negative log-likelihood (left) and the rate for behavior matched (right) for the average-participant cognitive model and the task solver for 2-step and 3-step conditions. Box plots indicate the median and interquartile range. Whiskers extend to the minimum and maximum values. Each connected pair of dots represents a single participant’s data. The task solver demonstrates significantly better performance.
Figure 3
Individuality transfer performance in Cross-Condition Transfer for the MDP task.
The plots compare the EIDT framework against an individualized cognitive model on negative log-likelihood (left) and rate for behavior matched (right) for both 2-step to 3-step and 3-step to 2-step transfer. Box plots indicate the median and interquartile range. Whiskers extend to the minimum and maximum values. Each connected pair of dots represents a single participant’s data. The EIDT model shows superior prediction accuracy.
Figure 4
Prediction performances as functions of latent space distance in the MDP task.
This cross-individual analysis shows the result of using a task solver generated from one participant to predict the behavior of another participant. The horizontal axis is the Euclidean distance between the latent representation of the two participants. The vertical axis shows the negative log-likelihood (left) and rate for behavior matched (right). Each dot represents one participant pair. Performance degrades as the distance between individuals increases, with the solid line showing the GLM fit. (A) 3-step to 2-step transfer. (B) 2-step to 3-step transfer.
Figure 5
Comparison of on-policy behavior between humans and EIDT-generated task solvers.
Each dot represents the performance of a single human participant (horizontal axis) versus their corresponding model (vertical axis) for one block. Plots show the total reward (left) and the rate of highly-rewarding action selected (right). (A) 3-step to 2-step transfer. (B) 2-step to 3-step transfer.
Figure 6
Mapping of Q-learning parameters to the individual latent space for the 3-step MDP task.
Each plot shows one dimension of the latent representation ( (left) or (right)) as a function of either the learning rate (, A) or the inverse temperature (, B) of simulated Q-learning agents. Black dots represent the latent representation produced by the encoder from the agent’s behavior. Blue dots show the fit from a GLM.
Figure 7
Task performance (rate of correct responses) in Within-Condition Prediction for the MNIST tasks.
Box plots indicate the median and interquartile range. Whiskers extend to the minimum and maximum values. Performance is compared across human participants, the RTNet model, and our task solver for the four experimental conditions (EA, ES, DA, and DS). All three show similar performance patterns.
Figure 8
Comparison of prediction performance in Within-Condition Prediction for the MNIST task.
The plots show the negative log-likelihood (left) and the rate for behavior matched (right) for the RTNet model and our task solver. Each connected pair of dots represents a single participant’s data. Box plots indicate the median and interquartile range. Whiskers extend to the minimum and maximum values. The task solver achieves significantly better prediction accuracy.
Figure 9
Individuality transfer performance in Cross-Condition Transfer for the MNIST task.
The plots compare the EIDT framework against the task solver (source) baseline across all 12 transfer directions on negative log-likelihood (top) and rate for behavior matched (bottom). Each connected pair of dots represents a single participant’s data. Box plots indicate the median and interquartile range. Whiskers extend to the minimum and maximum values. EIDT consistently demonstrates superior prediction accuracy.
Figure 10
EIDT captures individual-specific error patterns in the MNIST task.
The plots show the percentage of correct responses for each digit for four representative participants (blue bars) and their corresponding EIDT-generated models (gray bars). Data shown is for the ES target condition, with transfer from EA.
Figure 11
Prediction performance as a function of latent space distance in the MNIST task (transfer direction EA→DA).
This cross-individual analysis shows the result of using a task solver generated from one participant to predict the behavior of another participant. The horizontal axis is the Euclidean distance between the latent representation of the two participants. The vertical axis shows the negative log-likelihood (left) and rate for behavior matched (right). Each dot represents one participant pair. Performance degrades as the distance between individuals increases, with the solid line showing the GLM fit.
Figure 12
The 3-step MDP task.
(A) Tree diagram illustrating state-action transitions. (B) Flow of a single episode in the behavioral experiment for human participants.
Appendix 1—figure 1
Histograms of Q-learning parameters estimated from human participants’ behaviors in the MDP tasks.
The distributions for the learning rate (A) and inverse temperature (B) show considerable inter-individual variability, whereas the discount rate (C) and initial Q-value (D) are relatively consistent across participants.
Appendix 1—figure 2
Representative training and validation curves for the EIDT model in the MDP task.
The plots show the negative log-likelihood loss over training epochs for (A) 3-step to 2-step transfer and (B) 2-step to 3-step transfer. The star marker indicates the point of early stopping, where the validation loss was minimal.
Appendix 1—figure 3
Individual latent representations for the MDP task.
The plots show the two-dimensional latent space derived from behaviors in (A) the 3-step task and (B) the 2-step task. Square markers represent human participants, and dot markers represent simulated Q-learning agents.
Appendix 1—figure 4
Mapping of Q-learning parameters to the individual latent space for the 2-step MDP task.
Each plot shows one dimension of the latent representation ( or ) as a function of either the learning rate (, left) or the inverse temperature (, right) of simulated Q-learning agents. Black dots represent the latent representation from the agent’s behavior, while blue dots show the GLM fit.
Appendix 1—figure 5
Prediction performances for Q-learning agents as a function of latent space distance.
The plots show negative log-likelihood (left) and rate for behavior matched (right) in a cross-individual scenario. (A) 3-step to 2-step transfer. (B) 2-step to 3-step transfer.
Appendix 1—figure 6
Representatives training and validation curves for EIDT models in the MNIST task for each of the 12 transfer directions.
Training was stopped at the epoch with the minimum validation loss, indicated by the start marker.
Appendix 1—figure 7
Individual latent representations derived from human participants’ behaviors in the MNIST task.
Each panel shows the two-dimensional latent space generated when using a different experimental condition as the source.
Appendix 1—figure 8
Prediction performance (negative log-likelihood) as a function of latent space distance in the MNIST task.
Each panel shows the results for one of the 12 transfer directions. The negative log-likelihood (vertical axis) increases as the distance between the source and target individuals’ latent representations (horizontal axis) increases, indicating worse prediction performance. The solid line is the GLM fit.
Appendix 1—figure 9
Prediction performance (rate for behavior matched) as a function of latent space distance in the MNIST task.
Each panel shows the results for one of the 12 transfer directions. The rate for behavior matched (vertical axis) decreases as the distance between individuals’ latent representations (horizontal axis) increases. The solid line is the GLM fit.
Tables
Appendix 1—table 1
GLM fitting coefficients for the relationship between Q-learning parameters and the individual latent representation.
An asterisk (*) denotes statistical significance ().
| Coefficients | |||||
|---|---|---|---|---|---|
| Source | Variable | (bias) | () | () | () |
| 2-step | −0.613* | 0.096* | 0.064* | −0.035* | |
| 0.634* | −0.079* | −0.066* | 0.028 | ||
| 3-step | 1.157* | −0.388* | −0.121* | −0.148* | |
| −0.646* | 0.183* | 0.061* | −0.071* | ||
Appendix 1—table 2
GLM fitting coefficients () for the effect of latent space distance on prediction performance in the MNIST task.
All coefficients are statistically significant ().
| Negative log-likelihood | |||||
|---|---|---|---|---|---|
| Target | |||||
| EA | ES | DA | DS | ||
| Source | EA | — | 0.202 | 0.102 | 0.110 |
| ES | 0.117 | — | 0.185 | 0.111 | |
| DA | 0.199 | 0.158 | — | 0.076 | |
| DS | 0.201 | 0.186 | 0.174 | — | |
| Rate for behavior matched | |||||
| Target | |||||
| EA | ES | DA | DS | ||
| Source | EA | — | −0.142 | −0.103 | −0.187 |
| ES | −0.053 | — | −0.193 | −0.214 | |
| DA | −0.076 | −0.115 | — | −0.118 | |
| DS | −0.086 | −0.270 | −0.185 | — | |
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Predicting human decision-making across task conditions via individuality transfer
eLife 14:RP107163.
https://doi.org/10.7554/eLife.107163.3
