Neuromodulatory supervised learning

Biological nervous systems exist in an environment in which chemicals - neuromodulators - modify their function, notably the strengths of the neuron connections. This allows global modulation of the entire network by a parameter which, in artificial neuroendocrine systems, is typically an artificial "hormone'' decoupled from the network itself. Such systems have shown promise in implementing adaptive behaviour, particularly homeostasis. However, current implementations typically have their parameters set by hand, use pre-trained networks modulating only the output layer, or use complex and limited learning rules or evolutionary algorithms. The network described in this thesis is a simplified Neal/Timmis artificial neuroendocrine system with a single hidden layer, named UESMANN (Uniformly Excitatory Switching Modulatory Artificial Neural Network.) It is trained using a modified back-propagation of errors, with two sets of examples: one for each extremum of the modulator. Thus it uses a supervised learning algorithm. The work is divided into four parts, each consisting of a number of chapters. Part 1 consists of an introduction and outline of the thesis including the motivation behind it and a methodology, and an extensive literature review with a particular focus on sub-symbolic biologically-inspired adaptive systems. This is followed by Part Two, which introduces and describes the UESMANN network and performs an Monte Carlo analysis of random UESMANN nodes and networks performing pairings of boolean functions. This shows that a subset of boolean pairings is possible in a single node, but that all possible pairings of binary boolean functions can be performed with two hidden nodes: the same as the minimum required for a single boolean in a standard multi-layer perceptron. The modified back-propagation algorithm is then developed and tested on the boolean pairings with good results, with the resulting networks analysed in depth. Part Three shows that the network is capable of learning to transition between two line recognition functions (vertical and horizontal) and to transition between nominal and alternate labellings of the MNIST handwriting recognition database, with comparable performance with two other modulatory paradigms. In Part Four the network performed comparably with other techniques in a robot homeostasis task both in simulation and reality, showing particularly interesting transitional behaviour. The final part, Part Five, is a conclusion to the thesis as a whole with reference to the initial hypothesis and research questions, and describes further work which should be undertaken.