Using a combination of reinforcement learning techniques to create an autonomous general video game playing artificial intelligence.

Overview

The goal of this research task is to create an autonomous agent that can learn and play a variety of unseen video games. It is well established that common test beds for new machine learning algorithms are video game environments. Simple algorithms targeted to specific games can already produce masterful computer players. However, it is only now that we are beginning to see general algorithms being able to solve a variety of different genres and complexities. This area of research is of great interest at the moment with large corporations running regular competitions attempting to progress and solve this problem of generally intelligent video game agents. Both Microsoft (2018) and The GVG-AI Competition (Perez 2018) sponsored by DeepMind are regularly running competitions on general video game playing highlighting the large interest that corporations are showing in this research. Additionally, OpenAI (2018), another large AI company, provide a number of environments, including over 50 old Atari games to assist researchers in solving different AI problems. The great interest by large corporations and the availability of tools for researchers makes this an important and feasible research task.

The Problem

To design and develop an agent that learns and plays a variety of video games without prior knowledge of the rules or consequences of actions that the player can undertake. It will focus on single player games and attempt to create optimally acting or well performing agents according to the objectives of the game. It will use OpenAI’s Gym environment taking advantage of the over 50 different Atari games available on the platform (OpenAI 2018).

Team