Recent research conducted by the Cultural General Intelligence Team has revealed a breakthrough in the field of artificial agents. These agents are trained to mimic and retain the knowledge of both robots and humans. Utilizing deep reinforcement learning, the agents can extract and recall navigational expertise demonstrated by experts in real time, applying it across a wide range of new tasks.
The agents are trained and tested in procedurally generated 3D environments, encountering colorful goals within obstacle-filled terrains. A privileged “bot” with the correct goal sequence provides culturally transmitted information to the agents, demonstrating the skills necessary for successful navigation.
Through ablations, the research team identified a set of training ingredients essential for enabling cultural transmission. This new method called MEDAL-ADR, which includes memory, expert dropout, attentional bias towards the expert, and automatic domain randomization, outperformed other methods across various challenging tasks. Furthermore, the agents exhibited remarkable recall abilities, retaining demonstrated knowledge long after the experts had left.
Ultimately, the research paves the way for the evolution of more intelligent artificial agents, capable of flexible, high-recall, real-time cultural transmission without the use of human training data. The full paper detailing the team’s work can be accessed through this link: https://arxiv.org/abs/2203.00715