A Python/numpy implementation of SR-Dyna from [1] by Russek, Momennejad, Botvinick, Gershman and Daw, 2017.
Russek's Matlab code and Momennejad's Python tutorials were used for reference to fill in gaps when not fully specified.
Install requirements pip install -r requirements.txt, then run one of the task jupyter notebooks jupyter notebook ./sr-dyna-latent-learning.ipynb. Output animations are saved to ./out.
Main gridworld and algorithm code is in srdyna.py. Each notebook runs simulations attempted to match experiments presented in [1].
- Russek, E. M., Momennejad, I., Botvinick, M. M., Gershman, S. J., & Daw, N. D. (2017). Predictive representations can link model-based reinforcement learning to model-free mechanisms. PLoS computational biology, 13(9), e1005768.