Entity-Learning with Deep Reinforcement Learning: A Study on Different Abstractions of Input
Deep reinforcement learning is a method that introduced promising solutions to various setbacks that reinforcement learning historically presented. These advances are subject of many recent studies. They regard, for example, the possibility of automatically abstracting relevant information from the environment, and performing consistently using higher dimensional data from complex environments, like entities present in such environment. For this work, the ViZDoom platform will be used in order to compare the learning efficiency of distinct input proposals: one learning from raw image inputs, one learning from edited image inputs and the other using more structured data, identifying entities in the scene. The curriculum learning method adopted for training will gradually increase the complexity of the environment, validating the performance of the algorithm at each step. This work aims to contribute with a comparative study on the learning efficiency displayed by agents when presented to data with different abstraction-levels, learning from a dynamic environment over ever increasing fundamental skills.