Animal-AI Olympics Introduction Goals Dates Organisers

steampunkFOUR

Image Credit Squidoodle

The Animal-AI Olympics

Releasing into the wild April 2019

$10,000+ Prize Pool

AI has made significant progress in recent years, reaching superhuman performance on a wide range of tasks. Humans are no longer the best Go players, quiz-show contestants, or even, in some respects, the best doctors. Yet state-of-the art AI cannot compete with simple animals at adapting to unexpected changes in the environment. This competition pits our best AI approaches against the animal kingdom to determine if the great successes of AI are now ready to compete with the great successes of evolution at their own game.

We are proposing a new kind of AI competition. Instead of providing a specific task, we will provide a well-defined arena and a list of cognitive abilities that we will test for in that arena. Many elements will be fixed and known in advance. The tests will all use the same agent with the same inputs and actions. The goal will always be to retrieve the same food items by interacting with previously seen objects. However, the exact layout and variations of the tests will not be released until after the competition.

We expect this to be hard challenge. Winning this competition will require an AI system that can behave robustly and generalise to unseen cases. A perfect score will require a breakthrough in AI, well beyond current capabilities. However, even small successes will show that it is possible, not just to find useful patterns in data, but to extrapolate from these to an understanding of how the world works.

Competition Goals

steampunkFOUR

Experimental environments will be created in the Unity ML-Agents Toolkit, and the specification of the building blocks for all tasks will be freely available to all participants of the competition. Participants are allowed to use any methods and to experiment as much as they like in preparation for the competition, but the exact details of the tests will be kept secret. Participant performance will be measured on a range of tasks from simple combinations of the building blocks to complex tasks designed, each designed to probe certain cognitive abilities. Top prize will be for performance across the range of tests.

Timeline

Organising Committee


profilepic

Dr. Matthew Crosby
Imperial College London
Leverhulme CFI Postoctoral Researcher

profilepic

Benjamin Beyret
Imperial College London
Leverhulme CFI Research Assistant

profilepic

Prof. Murray Shanahan
DeepMind
Imperial College London,
Leverhulme CFI Spoke Leader


profilepic

Dr. Marta Halina
University of Cambridge
Leverhulme CFI Project Leader

profilepic

Dr. Lucy Cheke
University of Cambridge
Leverhulme CFI Associate Fellow

profilepic

Prof. José Hernández-Orallo
Universitat Politècnica de València
Leverhulme CFI Associate Fellow

External Collaborators

profilepic

Marek Rosa
CTO and CEO of GoodAI

profilepic

Olga Afanasjeva
COO of GoodAI

profilepic

Marek Havrda
Strategy Advisor GoodAI

Contact Address

m.crosby@imperial.ac.uk