Overcooked
Project Overview
- Duration: February 2020 - Present
- My role: Research assistant working alongside Prof. Anca Dragan and Dr. Andrew Critch
- Tools & Frameworks: Python, TensorFlow, matplotlib
Motivation
In today's world, an effective way to win a two-player game is to train an agent to play against other AI agents
that often take the form of past versions of itself. Many of these current algorithms are extremely effective against humans
in these competitive based games where only one person can win. However, not all settings are competitive and arguably, one of
the main goals of AI is to eventually create agents that can collaborate with humans. The difficulty is that modern algorithms for
competitive games do not produce equally strong results when used in a collaborative setting.
Thus, we are attempting to utilize a simplified version of the popular game Overcooked to model human-robot collaboration through cooperative
inverse reinforcement learning. I have created experiments to vary human and robot rationality to see how the interaction between the agents
changes as a result.