GENEREIT: generating multi-talented reinforcement learning agents

Abstract

AbstractCreating an intelligent system that is able to generalize and reach human or above-human performance in a variety of tasks will be part of the crowning achievement of Artificial General Intelligence. However, even though many steps have been taken towards this direction, they have critical shortcomings that prevent the research community from drawing a clear path towards that goal, such as limited learning capacity of a model, sample-inefficiency or low overall performance. In this paper, we propose GENEREIT, a meta-Reinforcement Learning model in which a single Deep Reinforcement Learning agent (meta-learner) is able to produce high-performance agents (inner-learners) for solving different environments under a single training session, in a sample-efficient way, as shown by primary results in a set of various toy-like environments. This is partially due to the fixed subset selection strategy implementation that allows the meta-learner to focus on tuning specific traits of the generated agents rather than tuning them completely. This, combined with our system’s modular design for introducing higher levels in the meta-learning hierarchy, can also be potentially immune to catastrophic forgetting and provide ample learning capacity.