We’re excited to announce the W&B SpaceInvaders Challenge, a reinforcement learning competition. Your goal is to train reinforcement learning agents in OpenAI's gym environment. The contestants with the top 3 scores will receive prizes and be invited to share their solutions with the community. The challenge is open to Qualcomm employees.
To enter the competition:
- Sign up for W&B using your Qualcomm email address
- Download the baseline notebook to get started. PLEASE NOTE: This notebook contains code for loading the gym environment, preprocessing data, calculating & logging the cumulative average reward metrics and saving your model files.
- Evaluate your model using the evaluation notebook.
- Submit your test runs here. See submissions instructions below.
New to online contests with W&B, reinforcement learning and/or OpenAI gym? No problem! We have posted resources to help you understand the W&B Python libraries, supported frameworks, suitable algorithms and some articles on neural networks below under the Resources section.
Questions? Use the #qualcomm-competition slack channel, or email contest@wandb.com.
Working with colleagues on this competition is encouraged! For teams we suggest sharing a Google Colab notebook to iterate on your training code. Once you're ready to submit your model, be sure to mention any other W&B usernames that were a part of your team in the notes section (What makes this run special?):
Google Colab is a convenient hosted environment you can use to run the baseline and iterate on your models quickly. To get started:
- Open the baseline notebook.
- Click "Open in playground" or "copy to drive" to create a copy of this notebook for yourself.
- Save a copy in Google Drive for yourself.
- To enable a GPU, please click Edit > Notebook Settings. Change the "hardware accelerator" to GPU.
- Step through each section, pressing play on the code blocks to run the cells.
- Add your own model code.
- Evaluate your model using the evaluation notebook.
- Submit your model (see instructions below).
You may submit your entries here. You'll need a Weights & Biases account to make submissions.
Each run must include the following files:
- Model file generated by wandb.save()
- Model training script (.py script or notebook)
- Any other files necessary to recreate the model
Please ensure that you log your model file (colab notebook, .py scripts etc) and all files necessary to recreate the model in your run using wandb.save(). Without this, we will be unable to evaluate your model. If you have trouble saving your model files to your wandb run, please email them to contest@wandb.com along with your wandb run url.
Also please ensure that your code is not in a public repo, but is visible to us by adding 'lavanyashukla' as a collaborator to your repo. We will use the model saved in the submitted run to recreate the model and evaluate it across the 5 random seeds.
Your objective is to maximize the best 100-episode average reward. This means your model will play the game for 100 episodes, and we will calculate a running average of the cumulative reward gained as each of the episodes is played. After 100 episodes, this cumulative running average will be your final score for the run.
We encourage you to submit as many runs as you like. To verify results, we will pick the top 5-10 submissions as ranked by the evaluation metric (best 100-episode average reward), and run these agents through the SpaceInvaders environment. We will evaluate how the agents do across 5 randomly generated seeds. This means, your agent will be run for 100 episodes with 5 different seeds and generate a best 100-episode average reward for each seed. We will take the average of these scores to get the final best 100-episode average reward.
Entries will be ranked from highest to lowest by the best 100-episode average reward received across the 5 seeds.
The Best W&B Report will be awarded to the individual who creates the best writeup explaining their model architecture, training process and results achieved. You can find sample W&B reports here, here and here.
- First place - $1000 travel gift card
- Second place - $500 travel gift card
- Third place - $200 travel gift card
- Best W&B Report - $500 travel gift card
- February 12 - Competition Announced
- April 30 - Deadline for final submissions
- June 1 - Winners announced
- TBD - Contest Retrospective Webinar
The deadlines are at 11:59PM on the days mentioned above PST.
You can find the leaderboard here.
- You are free to use any framework you feel comfortable in and submit as many times a day as you wish.
- We don’t allow the use of automated machine learning tool(s) in this competition.
- You may only have one account per individual.
- You can share small snippets of the code online or in our slack community, but not the full solution - until the submission deadline has passed.
- You can submit as many runs as you like.
- Weights & Biases Docs
- Playing Atari with Deep Reinforcement Learning
- How RL agents learn to play Atari games
- Overcoming catastrophic forgetting in neural nets - Paper
- Frame Skipping and Pre-Processing for Deep Q-Networks on Atari 2600 Games
- The A3C Algorithm
- The Deep Q-learning algorithm
Weights & Biases is an experiment tracking platform for deep learning. Import our python package into any training script with a few lines of code. Explore how hyperparameters affect model performance in realtime, record and visualize every detail of your research, compare metrics across thousands of runs, reproduce and quickly share findings with collaborators.