https://github.com/mnoukhov/ecn

Experiments on Emergent Communication Through Negotiation
https://github.com/mnoukhov/ecn

Last synced: 19 days ago
JSON representation

Experiments on Emergent Communication Through Negotiation

• Host: GitHub
• URL: https://github.com/mnoukhov/ecn
• Owner: mnoukhov
• License: mit
• Created: 2018-04-19T01:57:46.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2023-05-04T02:08:34.000Z (over 1 year ago)
• Last Synced: 2024-10-28T00:21:27.668Z (2 months ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 3.18 MB
• Stars: 1
• Watchers: 2
• Forks: 1
• Open Issues: 9
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Emergent Communication under Competition (AAMAS 2021) experiments on the negotiation game

Experiments from the paper [Emergent Communication under Competition (AAMAS 2021)](https://arxiv.org/abs/2101.10276) on the negotiation game setup of [Emergent Communication through Negotation (Cao et al, 2018)](https://arxiv.org/pdf/1804.03980.pdf). Code was initially built off of [Hugh Perkins' repo](https://github.com/ASAPPinc/emergent_comms_negotiation)

For code relating to the biased circular sender-receiver game see [here](https://github.com/mnoukhov/emergent-compete)

## Setup

### Requirements

I recommend running in a docker container which you can build with the `Dockerfile`.

Otherwise, use a python `venv` and install the requirements

```

python -m venv env

source env/bin/activate

pip install -r requirement.txt

```

Optionally you can also set up `PYTHONPATH` variables to facilitate imports. Assuming you cloned the repo into `$HOME/ecn` you can use

```

export PROJECT=$HOME/ecn

export PYTHONPATH=$PYTHONPATH:$PROJECT

```

### Logging

I use `wandb` for logging and recommend it! You can activate logging to it with the `--wandb` arg.

Otherwise we log in `json` format to a file in `--logdir` specified as `{args.name}_{slurmid}_{timestamp}`. `slurmid` is useful if you're running on a cluster with `slurm` and it is pulled automatically

## Run

Use the scripts in `/scripts` to run the code. Each script name corresponds to the setup it runs

- `self-*` runs selfish agents

- `*-none-*` uses no communication channels

- `*-ling-*` uses only the linguistic communication channel

- `*-prop-*` uses only the proposal communication channel

- `*-both-*` uses both the proposal and linguistic communication channels

- `*-masking` uses the masked linguistic channel, where agents learn a mask over their proposals and their linguistic utterance is simply the mask multiplied with their proposal

If you want change a flag you can add it to the end of the command, e.g. entropy regularization of term policy

```

python scripts/self-none.py --term-entropy-reg 0.1

```

The scripts define the hyperparameters for each experiment

## Reproducing Graphs

We run 5 seeds of every experiment and the plot the mean and 95% confidence interval over the seeds. Run the script with `--seed` for 5 different values (e.g. `0-4`) and log to `wandb`

From `wandb`, download the `test_reward` for all 5 seeds into a `csv` for the corresponding experiment. Plot it using `notebooks/plot_wandb.ipynb`

If you're using logfiles instead of `wandb` use `notebooks/plot_logfile.ipynb`