Abstract:
Artificial Intelligence (AI) are machines designed to think and behave as humans would.
Taking AI and placing it into a virtual world they become known as AI agents which uses
the knowledge it gained from training to perform tasks in the world. AI agents in the
virtual world has only been able to perform a narrow set of tasks with specialised models
in environments with limited complexity and diversity. A rich world that requires an
agent to continuously learn from and adapt to a wide variety of open-ended tasks and use
previously gained knowledge to determine the next course of action will render the agent
incapable. In order to investigate the AI teaching methods applied to instruct the agent to
perform basic tasks in Minecraft in order to identify which AI teaching methods will yield
the best results, a systematic literature review was conducted by extracting 57 papers and
identifying themes and sub-themes that suited AI agent training methods and functions.
This was to discover wat AI training methods can be implemented to enable an agent to
perform tasks in a complex and rich world, contributing to game-based learning. The
study found that a well-integrated Reinforcement Learning (RL) method with an effective
reward system equipped the agent with the necessary knowledge to be able to perform
tasks on a more complex level. A list of unique methods was integrated with RL such as
Newtonian Action Advice (NAA), Behavioural Cloning (BC), VideoPreTraining (VPT),
human demonstrations, and natural language commands to achieve a certain goal. This
meant that AI agents can be taught to perform open ended tasks in a complex environment
by setting up a well thought out framework on how to teach the agent in various areas
leading to the possibility to incorporate those teachings into the real world through gamebased
learning.
