DeepHack RL consists of 3 stages.

Stage 1. Selection

Team members apply for participation in hackathon by filling a form. Solution for Atari game #1 Skiing is uploaded on OpenAI Gym from github account associated with the team. The solution will be evaluated for the full skier run till the the bottom of the ski course. Code for the uploaded solution should be sent to info@deephack.me from an e-mail associated with the team. Solution submission deadline is 23:59 (Moscow time) 26 Jan 2017. TOP-50 of registered applicants, who has the highest score at OpenAI Gym leaderboard of Atari Skiing game as an individual or a team member will be invited to participate in hackathon.

Stage 2. On site competition

Atari game #2 will be announced on 6 Feb 2017. Solutions will be evaluated via OpenAI Gym leaderboard. Code for the uploaded solution should be sent to info@deephack.me from an e-mail associated with the team.

Stage 3. Finals

Atari game #3 will be announced on 10 Feb 2017. For finals, teams should provide video of gameplay for all 3 games till the morning of 12 Feb 2017. Finals will be held as a single-elimination tournament (playoff). For every match in a tournament one of the 3 games will be selected randomly.

Solutions should meet the following requirements.

1. The same agent (algorithm) should be used to learn and generate solutions for all 3 games in finals. The agent should use only information provided by the OpenAI gym environment's "step" function for training and generation of solution. Environment specific handcrafted features and rewards are prohibited.
2. The agent policy should not be fitted to specific seed of random generator. Method 'env.seed' should not be used during testing of the agent
3. The agent should not be trained on the human gameplay or initialized with parameters of the third party solution.
   
4. The agent should be trained only on games as they are implemented on OpenAI Gym or ALE. During training environments should be initialized with any parameter settings provided here - https://github.com/openai/gym/blob/master/gym/envs/__init__.py#L248:L330. Number of time steps in an environment can be set to any value during training of the agent.
5. Final solution is generated without interruption of gameplay. Game saving and alternative replays are not allowed during generation of final solutions. This restriction doesn't apply to the training of the agent.
   
6. There are no restriction on the means of agent implementation, any learning algorithms and architectures can be used.
   
7. Code for generation of final solutions should be open before the finals on 12 Feb 2017.
8. Code for generation of final solutions should not be modified after Atari game #3 announcement. The only exception is dimensionality of actions space.

Yes. You can find the baseline solution here - https://github.com/openai/universe-starter-agent.

Games selected for competition will have less than 100% normalized score according to the "Table 5. Normalized scores across all games. Starting with Human Starts." from the paper Wang, Ziyu, Nando de Freitas, and Marc Lanctot. "Dueling network architectures for deep reinforcement learning." arXiv preprint arXiv:1511.06581 (2015). https://arxiv.org/abs/1511.06581

Of course, you may register and participate as a team of size 1 but as our previous experience indicates best results are achieved by teams of size 3-7 people. We will hold a teambuilding event before starting the hackathon. The point of the hackathon is to emphasise teamwork.

No, there is not.

School participants are welcome to stay at MIPT dormitory. The rate is about $10 per night per person and it is expected to be covered by participant. Coffe-breaks will be served along the school. Considering travel expenses (visa, flight or train tickets) - we will be ready to answer this question later. Please, give us all relevant information by the registration form.

The Moscow Institute of Physics and Technology will host the event and provide accomodation for all participants - https://mipt.ru/en/

Yes. If you want to be the free listener, please give us this information during the registration. Free listeners are not eligible to participate in the competition.

Yes. We plan to give each team a 24/7 access to remote instance with GPU.

Poster session is substituted with extreme coding.