GENEREIT: generating multi-talented reinforcement learning agents

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