[robotics-worldwide][jobs] PhD position in RL for humanoid robots

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view

[robotics-worldwide][jobs] PhD position in RL for humanoid robots

We are looking for PhD student in reinforcement learning for humanoid robots (Topic 2:  Learning and Planning with Emergent Hierarchical Representations and Decaying Short-Term Memory) at the Centre for Robotics and Neural Systems at the University of Plymouth UK.

A brief description of the project is included below.
The application deadline is May 29th, 12noon, UK time.
The project starts October 1st.

Further details on the project, funding and application procedure are available at:

Prospective candidates are invited to contact the director of studies directly.

Best regards,
Torbjorn Dahl
Torbjørn S. DAHL, MEng, ACGI, PhD
Lecturer in Software Engineering
Centre for Robotics and Neural Systems
School of Computing and Mathematics, Plymouth University
Drake Circus, Plymouth PL4 8AA, UK
web-site: http://www.tech.plym.ac.uk/SoCCE/crns/staff/dahl/


This project will develop a new generation of artificial neural networks for reinforcement learning (RL), significantly increasing the applicability of RL in areas such as robotics.  The new algorithms will make use of two central features of biological memory to limit their space and time requirements; emergent hierarchical representations and decaying short-term memory (STM). The algorithms developed will use hierarchical memory structures that grow according to a given RL problem.  The hierarchical structures will be made from multiple Kohonen networks and will provide the algorithms’ long-term memory (LTM) capabilities.  The Kohonen networks will be augmented with decaying node activation values providing explicit STM capabilities within a connectionist framework.  The main encoding mechanism in the new algorithms uses the decaying activation values (STM) of one network layer as input to update the weights (LTM) of other layers.  This architecture has allowed us to develop new, more efficient, mechanisms for key RL functions including hidden state identification and future reward estimation.  We have already demonstrated some of these mechanisms using analytical methods and lossless memory encoding (Pierris and Dahl, 2014).  This project will build on those results to produce true connectionist algorithms that can be parallelised, e.g., using Cuda technology.  The work will consider traditional RL benchmarks as well as robot learning problems and will use the University of Plymouth’s Nao or iCub humanoid robots.

Related reference

Pierris G. and Dahl T. S., Humanoid Tactile Gesture Production using a Hierarchical SOM-based Encoding. IEEE Transactions on Autonomous Mental Development, 6(2):153-167, 2014.