MSc graduation project: Solving the Cocktail Party problem with Intelligent Autonomous Agents

Context

BIASlab (Fig.1, http://biaslab.org, FLUX-7.060) is a subgroup of the Signal Processing Systems (SPS) that aims to develop Intelligent Autonomous Agents (IA). These agents interact with their environment through their sensors and actuators in order to learn purposeful behavior, e.g., to navigate, play soccer or they may learn to decode speech signals under bad acoustic conditions. Our research projects are inspired by the latest insights from machine learning, computational neuroscience and signal processing.

Project Description

In this project, you are challenged to design an agent that learns to solve the cocktail party problem through on-the-spot interactions with a (human) listener.

The cocktail party problem refers to the issue of not being able to understand your conversation partner in the presence of many simultaneously competing voices (Fig.2). The learning protocol is displayed in Fig.3. A listener wears earbuds that are capable to process audio signals in real-time (like hearing aids). In response to a detected problem, the agent proposes the most promising alternative parameter settings for the audio algorithm (the TRY step). Next, the new audio algorithm is executed in the ear buds and evaluated by the listener (EXECUTE and EVALUATE steps). Based on the listener’s appraisal, the agent should now update its model of the world (LEARN step). This design loop repeats in real-time until the listener indicates that the problem has been solved.

This project will get you involved with the latest artificial intelligence methods, since the agent needs to (1) learn from each interaction and (2) be smart about selecting the most promising algorithm candidates. It will also give you an opportunity to learn about how biological brains solve real-time design issues.

Timing

Start date: the project is available from September 2017 (or any time thereafter).

Duration: 9 months (fte).

Financial Support

The audio solutions company GN (http://www.gn.com/) may financially support strong candidates (qualifications to be discussed with prof. de Vries) by a GN scholarship.

Contact

For more information about this project, please contact Prof. Bert de Vries (bert.de.vries@tue.nl). Also, feel free to make an appointment to discuss alternative projects with intelligent autonomous agents that you might be interested in.