The American Association of Geographers Annual Conference will take place in Boston, April 5-9th 2017, and the call for papers has just gone out if any of you are interested in submitting http://www.aag.org/cs/annualmeeting/call_for_papers. This is the largest gathering of geographers and cartographers in the world and there are several special sessions on Geospatial data archiving and preservation, sponsored by the Archives Committee and, of course, hundreds and hundreds of papers from across the geographical and cartographic disciplines. ( http://www.aag.org/annualmeeting ). The archive papers might be of special interest across the Library has we are now the depository for the AAG archives.
This year I will be giving a Plenary talk called, “Finding Our Way in the World: the Neuro-Computational Foundations of the Cognitive Map” and be part of a special Presidents Session and Panel called:
Recently, geographers have taken up the question of robots and robotic technologies within the confines of a broadly engaged human and environmental geography. From the rise of robotic warfare to the development of smart cities and borders to the reliance on code, big data analytics, and autonomous sensing systems in environmental management, geographers are interrogating what robots and robotic technologies mean not only for discipline, surveillance, and security, but for making and remaking everyday life and the socio-natural environment.
My paper for the session deals with research at the edge of robotics and machine learning and with new concepts in computational spatial embedding and is called:
How far can they go?
Robots, Convolutional Neural Networks, and Model-Free Reinforcement Learning
Recent advancements in deep learning and convolutional neural network algorithms have transformed the tasks that can be performed by robotic agents in a wide variety of complex environments. Robots now more than ever use model-free learning methods in which the robotic agent needs to represent nothing in its programming about the environment in which it operates, but rather learns by interacting directly with it in probabilistic ways. Using these types of algorithms the agent’s goal is to maximize the total amount of reward it receives by optimizing what is known as a policy or value function. Examples of these kinds of robots include, the OBELIX Robot, the Zebra Zero Robot, the Sacros Humanoid DB and, most recently, a Go playing agent called Alphago that beat a world champion player for the first time early in 2016. These powerful kinds of adaptable and general learning frameworks call into question previous definitions of spatial interaction with the environment and constitute a new kind of environmental embedding in which the programmed model used by the robot agent is the environment itself. This paper gives an introduction to reinforcement learning algorithms and discusses the new kind of geographic and spatial embedding and embodiment that model-free dynamic programming methods represent. It further goes on to speculate on how these methods will accelerate the number of robotic agents interacting with humans in the near future across tasks once thought unapproachable by robot agents.
All the best. Hope to see some of you there.
John Hessler, FRGS
Specialist in Modern Cartography and Geographic Information Sciences
Geography and Map Division
Library of Congress
Full Bio: http://blogs.loc.gov/maps/author/jhes/