lakshmiaddepalli / human-priors-and-deep-reinforcement-learning-for_video-games

Computational Cognitive modelling, Reinforcement Learning NYU Spring 2020, Dr Brenden Lake and Dr. Todd Gureckis.

Human Priors and Deep Reinforcement Learning for Video Games

This is the project for Computational Cognitive Modeling Course NYU PSYCH-GA 3405.002 / DS-GA 1016 https://brendenlake.github.io/CCM-site/ under Professor Dr. Brenden Lake and Dr. Todd Gureckis.

Here we analyze on how having a prior knowledge is helpful for humans in playing video games and compares its game play with that of an Reinforcement Learning Agent. We try to answer the questions:

1. Are Humans better in solving complex video games than an RL trained agent?

2. Does having prior knowledge about the world help humans make better decisions to solve an complex game than an RL agent?

We consider the Flappy Bird Video Game, and conduct different experiments to get a quantitative aspect of how important having prior knowledgehelps in the performance of humans. We modify the environment on various basis, some being masking visual information or important information needed for efficient game-play and provide a comparison of results between human and an RL agent performance.

Keywords:Reinforcement Learning; Human Understanding; Exploration; Priors; Deep Q Learning; Dueling DQN; Convolutional Neural Networks; Semantics; Affordances; Cognitive Modelling; Flappy Bird