Instructions with Complex Control-Flow Entailing Machine Learning

Abstract

Reinforcement learning is when the system is allowed to make its own decisions based on what it learns. There are 2 types of observations, formative and summative. These observations have been identified as crucially important for neural network training of complicated tasks with conditional control flow. The central theme of this paper is applying reinforcement learning to follow instructions with complex control-flow. The authors study a special but important subset of multi- task reinforcement learning problems, namely instructions with complex control-flow in this work. They develop an encoding and attention architecture to achieve the research objective.