Human-AI Teaming in Games

People

Supervisor

Research areas

Description

Video games always include an element of automation, which is usually invisible to the player. The automation removes the burden of tedious and low-level tasks, freeing up the player to concentrate on more interesting and challenging parts of the game. For example, a player might order a unit to attack another unit but will not need to reissue this command repeatedly nor in detail. Instead, the unit will attack its target until it is defeated. As humans have certain perceptual, cognitive, and memory limitations, finding the right amount and right types of things for the player to do and attend to is fundamental to their initial and lasting engagement in a game. This principle can be extended to tasks and software more generally, but the focus changes when the goal is not primarily enjoyment, but also performance or decision-making. 

On the other hand, AI research in games has focused almost entirely on performance and developing human-like or human-superior AI. Although AI has performed well in traditional games with perfect information and relatively small state spaces (e.g., board games), there are many challenges and barriers when moving to complex video games with incomplete information, large state spaces, and sparse rewards and/or limited feedback. Games that are quite simple for humans to play are currently impossible for machine learning to master. Previous game AI research has also focused on individual, rather than team, performance or teamwork. 

If our overall goal is to facilitate superior performance and decision-making in real-world situations, it makes sense to combine the varying abilities of humans and machines. There are different ways to think about approaching this problem. Many games, and some research, try to create mixed teams of humans and AIs to varying effect. However, the shortfall is that each human and AI agent is expected to perform the entire spectrum of tasks. If we recognise that humans are superior at certain tasks and machines are superior at other tasks, we can look to develop a distributed collaboration more akin to the automation that exists to support players in most games. The questions that arise are: where do we draw the line to divide the tasks between humans and machines? What are the best ways for human-machine collaborations to communicate, share information, and make decisions? Can a hybrid human-AI collaboration outperform both AIs and humans alone, or more traditional models of mixed human-AI teams?

This project aims to:

  • Identify and define relative human and AI strengths and weaknesses in operating in video game environments with respect to performance and optimal decision-making.

  • Design, construct, and test AI models based on identified strengths and weaknesses.

  • Determine how humans and AIs can best collaborate and communicate with respect to their identified strengths and tasks.

  • Identify, design, develop, and test methods of information sharing, sense-making, and decision-making between humans and AI in video game environments.

This project involves investigating a specific approach to the above problem. If you would like to participate in this project, please send me a 1-2 paragraph proposal for your specific approach.

 

Keywords

AI, machine learning, human-AI interaction, games, HCI

Updated:  1 June 2019/Responsible Officer:  Dean, CECS/Page Contact:  CECS Marketing