ENKI: Cookie Monster Demo

Learn how the ENKI engine uses Machine Learning to predict future outcomes and adapt in real time.

React AI’s ENKI engine runs in real-time collecting sensor data and making decisions based on future predictions.

The ENKI engine controls the “Cookie Monster” using closed-loop analytics to guide it through its world by taking actions (back/forward, left/right) to maximise its rewards (eating cookies – the green circles). When initiated the engine has no information about its environment. Without the configuration, the rules, or the goal of the game, it has to learn how to operate to maximise reward.  More importantly, no information is supplied about messages it will receive, nor configuration of the fields. We simply tell it it to pick actions to maximise the score, and let it go from there.

Predictions using ENKI

What's under the React AI hood?

The ENKI engine is leveraging an innovative combination of several machine learning techniques (feed forward neural nets, recurrent neural nets, reinforcement learning and data fusion) to extract patterns from the message streams and build a holistic predictive model of the Cookie Monster’s world.

Data fusion is the integration of multiple varied sources of information to reduce noise and improve understanding. ENKI was designed to merge messages from different sources on different timescales (asynchonously), for example the Cookie Monster’s sense of touch, movement, and sight. Touch events happen irregularly, movement messages depend on the input of the ENKI engine or of the player, while the cookie monster has continuous sight of its environment. All these are combined into a single vector.

Closed Loop Analytics is a Machine Learning technique that relies on a live feedback loop between a Machine Learning engine and the real world. The ENKI engine suggests what it believes to be the best action for the agent to be obtain rewards –  ENKI then uses the feedback from testing these suggestions to learn and correct its model of the world.