Two weeks to go!
The deadline for the Round 1 submissions is August 14th 2017 (23:59 CET).
Make sure you do not miss it by preparing for the submission procedure in advance!
Overview of submission procedure
You are expected to send an email to email@example.com with basic information we request (name, contacts, competition category, evaluation hardware, target OS and sharing preference), and a white paper explaining the main principles and motivations behind the agent’s design in a brief, structured manner (2 pages max.; papers over two pages must include a one page abstract).
After that, if you compete for the quantitative prize, we will ask you to upload the data to Microsoft Azure Blob Storage.
We prepared a detailed step-by-step guide through the submission process on Linux and on Windows. You can also find the detailed requirements and submission email template there.
Challenge submissions fall into two categories: quantitative and qualitative.
The quantitative submissions will be evaluated on a private evaluation curriculum, a series of approximately 40 tasks of gradually increasing complexity. The agent will need to solve all the tasks from the curriculum, in the least number of steps.
The winning agent must learn in a gradual way: it will be tested for non-forgetting and reuse of skills on the tasks from the private evaluation curriculum.
In more detail: non-forgetting will be tested by re-running an agent on a previously solved task A, after it solved other tasks B, C, etc. The agent should solve task A in less simulation steps, or at least not slower than the first time it encountered it. Reuse of skills will be measured by comparing agent's performance on task B with and without prior knowledge from tasks that are prerequisite for task B.
The qualitative submissions will be white papers describing an architecture for gradual learning. They will be evaluated by jury, which will look for an idea or design that shows the best promise for scalable gradual learning.
More about the evaluation procedure can be found in the Specifications document.
Microsoft Azure access for participants
If you would like to train or test your agent on Microsoft Azure, you can get free access from our technological partner Microsoft Czech Republic and Slovakia. If you intend to use GPU-equipped instances, you need to make a specific request, even if you already have access to Azure.
Enabling access to GPU-equipped VMs can take several workdays. We recommend to ask for the access soon after you realize you or your team may need it.