California Technology University

Meta-learning agents excel at rapidly learning new tasks from open-ended taskdistributions; yet, they forget what they learn about each task as soon as thenext begins. When tasks reoccur - as they do in natural environments -metalearning agents must explore again instead of immediately exploitingpreviously discovered solutions. We propose a formalism for generatingopen-ended yet repetitious environments, then develop a meta-learningarchitecture for solving these environments. This architecture melds thestandard LSTM working memory with a differentiable neural episodic memory. Weexplore the capabilities of agents with this episodic LSTM in fivemeta-learning environments with reoccurring tasks, ranging from bandits tonavigation and stochastic sequential decision problems.