⚠ Note: This is the code for my article Meta-Reinforcement Learning on FloydHub. This repository is for the two-step task. For the Harlow task see this repository instead.⚠
We reproduced the simulations regarding the two-step task as described in the two papers:
- Learning to Reinforcement Learn, Wang et al., 2016
- Prefrontal cortex as a meta-reinforcement learning system, Wang et al., 2018
For a short explanation of the two-step task, see two-step-task.ipynb.
I answer questions and give more informations here:
We reproduced the plot from Prefrontal cortex as a meta-reinforcement learning system (Simulation 4, Figure b), on the right). We launched n=8 trainings using different seeds, but with the same hyperparameters as the paper, to compare to the results obtained by Wang et al.
For each seed, the training consisted of 20k episodes of 100 trials (instead of 10k episodes of 100 trials in the paper). The reason for our number of episodes choice is that, in our case, the learning seemed to converge after around ~20k episodes for most seeds.
After training, we tested the 8 different models for 300 further episodes (like in the paper), with the weights of the LSTM being fixed.
Here is the side by side comparison of our results (on the left) with the results from the paper (on the right):
Clone this repository and install the other dependencies with pip3:
git clone https://github.com/mtrazzi/two-step-task.git
cd two-step-task
pip3 install -U -r requirements.txt
The results of our attempt at reproducing the results of "Learning to Reinforcement Learn" as described in the arxiv pre-print are included in the notebook arxiv.ipynb.
For the implementation of the two-step task as descibed in "Prefrontal cortex as a meta-reinforcement learning system", we have included two different implementation:
- The first one, in
biorxiv-first-try.ipynbis an interpretation of the experiments where the LSTM is fed actions and rewards both from first and second stage. We saw that the convergence was much slower than in the paper, so we changed our approach. - The second implementation, in
biorxiv-final.ipynb, feeds only actions from first stage and rewards from second stage in the LSTM. Furthermore, the simulation was launched for 8 different seeds, to compare with the plot with eight seeds from the paper. We find that our results closely matched the results from the paper.
The notebooks are at the root of the repository.
Our trained models can be find in the directory results. In particular, our two trainings (with respectively 137k and 40k episodes) related to the arxiv pre-print "Learning to reinforcement learn" can be found in results/arxiv, and our two implementations for the biorxiv pre-print "Prefrontal cortex as a meta-reinforcement learning system" can be found in results/biorxiv.
Meta-RL
├── LICENSE
├── README.md
├── arxiv.ipynb
├── biorxiv-final.ipynb
├── biorxiv-first-try.ipynb
├── helper.py
├── requirements.txt
└── results
├── arxiv
│ ├── arxiv_137k
│ └── arxiv_40k
└── biorxiv
├── final
└── first-tryThe plots were generated with tensorboard.
For instance, to plot the reward for the eight simulations (as mentioned in "Main Results") do:
tensorboard --logdir=results/biorxiv/finalFor each sub-directory (i.e. arxiv/arxiv_137k, arxiv/arxiv_40k, biorxiv/final and biorxiv/first-try) we included the trained models.
- For
arxiv/arxiv_137k,arxiv/arxiv_40kandbiorxiv/first-try, the models can be found attrain/model_meta_context/model-[number of training episodes]/. - For
biorxiv/final, it's atmodel_[seed number]/model-20000/
To test the trained model, the load_modelvariable must be set to Trueand the load_model_path must be set to one of the path mentionned above (e.g. load_model_path=results/biorxiv/final/model_[seed number]/model-20000/).
Michaël Trazzi and Yasmine Hamdani, under the supervision of Olivier Sigaud.
This work uses awjuliani's Meta-RL implementation allowing only one thread/worker (so it was equivalent to a single-threaded A2C LSTM). We completely changed the code for the bandits to adapt to the two-step task, while keeping the same API.
All the code for the plots/gifs is ours.