Many neural control systems are at least roughly optimized, but how is optimal control
learned in the brain? There are algorithms for this purpose, but in their present forms they aren’t suited for biological neural networks because they rely on a type of communication that isn’t available in the brain, namely weight transport — transmitting the strengths, or “weights”, of individual synapses to other synapses and neurons. Here I show how optimal control can be learned without weight transport. I explore three complementary approaches. In the first, I show that the control-theory
concept of feedback linearization can form the basis for a simple mechanism that learns
roughly optimal control, at least in some sensorimotor tasks. Second, I describe a method based on Pontryagin’s Minimum Principle of optimal control, by which a network without weight transport might achieve optimal open-loop control. Third, I describe a mechanism for building optimal feedback controllers, without weight transport, by a method based on generalized Hamilton-Jacobi-Bellman equations. Finally, I argue that the issues raised in these three projects apply quite broadly, i.e. most control algorithms rely on weight transport in many different ways, but it may be possible to recast them into forms that are free of such transport by the mechanisms I propose.
| Identifer | oai:union.ndltd.org:LACETR/oai:collectionscanada.gc.ca:OTU.1807/24706 |
| Date | 05 August 2010 |
| Creators | Chinta Venkateswararao, Lakshminarayan |
| Contributors | Tweed, Douglas |
| Source Sets | Library and Archives Canada ETDs Repository / Centre d'archives des thèses électroniques de Bibliothèque et Archives Canada |
| Language | en_ca |
| Detected Language | English |
| Type | Thesis |
Page generated in 0.0021 seconds