The Learning Agent Triangle: Towards a Unified Disambiguation of the AGI Challenge.

In this short paper, we tackle an ambiguity issue in the discussion of how to build Artificial General Intelligence (AGI), with the goal of removing a communication barrier which is arguably slowing down its development. Due to the openness of the AGI problem, many design ideas describe some aspects of a learning agent but ignore or make implicit assumptions about other key features. We argue that, when sharing AGI design hypothesis, it is necessary to describe or constrain three specific key aspects of the agent, and we explain why only discussing about a subset of these aspects reduces the usefulness of the design hypothesis for the progress towards AGI. We disambiguate the design of a machine learning agent into what we call the Learning Agent Triangle, formed by the architecture, the objective goal and the optimization algorithm, which are conditioned by the computational resources. It must be noted that, even if the learning agent triangle might not be the most general or accurate way to describe any kind of agent, this model can be used as a framework to guide a description of an AGI in a complete enough way that the value of the contribution is not negatively affected by ambiguity or communication issues.