Projects:2017s2-220 Alternative Approaches to AI for the Soccer Table

From Projects
Jump to: navigation, search

Summary

The project aims at the development of an AI agent that can learn to play the table-top game Foosball in a simulated environment. Our objective is for the agent to be able to, on average, play better than a consistent, randomly moving opponent. Another objective is for the software integrating the simulation and agent training to be a useable training testbed. This project is a case-study application of the emerging field of Reinforced Deep Learning and attempts to replicate a Dueling Double Deep Q Network model by DeepMind.

The decision-making, deep-learning agent perceives its environment and takes actions that maximise its chance of success in the game. The agent is trained using Reinforcement Learning in a software simulation of a Foosball table. The simulated environment to which the agent interacts was implemented from a game’s source code.

The software package TensorFlow is used in the language Python, and the learning process is executed on high-performance computing system. The learning process allows the agent to develop its Foosball-playing ability.

When trained for 750,000 game frames, with a simplified game test-case, the agent clearly learns to intercept the ball. However, with a full game test-case, the agent does not show any obvious learnt intelligent behaviour. Likely due to a limitation in memory allocation to training and efficiency, the agent was not able to replicate the previous success of the model by DeepMind.

Project Team

Students:

Daniel Calandro

Liang Xu

Supervisors:

Dr Braden Phillips

Dr Hong Gunn Chew