What DeepMind’s AlphaZero did for Go, we are doing for the battlespace to train Warfighters.
DeUmbra develops intelligent simulated agents to help train and guide Warfighters. Training in simulation allows for a breadth of experience not possible with live training, and populating simulations with intelligent agents allows training to be done inexpensively because other humans need not be directly involved.
In addition, the intelligence of the agents allows Warfighters to experience adversaries with creative and novel behaviors, and the intelligent agents can also help guide the Warfighter by suggesting actions.
We are building our intelligent agents using methods from both deep neural networks and developmental learning. Developmental learning enables agents to build up competency incrementally, mirroring how human children learn more complex concepts on top of previously mastered ones. Our agents are developed in the same simulated environments used for Warfighter training, and they are taught to behave by being given increasingly difficult tasks. To achieve these tasks, they must learn to understand and manipulate the dynamics of the environment.