The Department of Computer Science at the University of Cyprus cordially invites you to the Colloquium entitled:
Why do we Commit to an Uncertain Future?
Speaker: Dr. Gaye Banfield
Empirical data in psychology suggests that we recognize we have self-control problems and attempt to overcome them by exercising precommmitment, which biases our future choices to a larger, later reward. The behavioral model of self-control as an internal process is taken from psychology and implemented, using a top-down approach, as a computational model of the human brain. The higher and lower brain systems, represented by two Artificial Neural Networks (ANNs) using reinforcement learning, are viewed as cooperating for the benefit of the organism, as opposed to the classical view of the higher brain overriding the lower brain. The ANNs are implemented as two players, learning simultaneously, but independently. Psychological studies suggest that the structure of the self-control problem can be likened to the Iterated Prisoner's Dilemma game in that cooperation is to defection what self-control is to impulsiveness. I hypothesise that increasing precommitment increases the probability of cooperating with oneself in the future. To this aim, precommitment is implemented in one of three ways. The first investigates the effect of implementing precommitment by simply varying the input value of the ANN's bias node between 0 and 1 (instead of fixed as 1). This method is referred to as the 'variable bias' method. The second implements precommitment as an extra input to the ANNs in the 2-ANNs model. In this case the ANN's threshold is implemented in the usual way, i.e., as a node with an input value of 1 whose weight is trainable in the same way as the other nodes in the network and precommitment is implemented as an additional node to the input layer. This method is referred to as the 'extra input bias' method. The final method implements a bias towards future rewards as a differential bias applied to the payoff matrix. Again the ANN's threshold is implemented in the usual way, i.e., as a node with an input value of 1 whose weight is trainable in the same way as the other nodes in the network. This method is referred to as the 'differential bias method'. Finally, I investigate what role evolution has played in shaping our willingness to precommit to future rewards by subjecting the model to simulation of evolutionary adaptation. Results suggest that evolution, as opposed to learning is the key player.
Dr Gaye Banfield has been involved in computing in one form or another for twenty-six years. She was awarded a Bachelor of Science from the University of Queensland in 1983 majoring in Computer Science with electives in Mathematics and Psychology. From 1983 to 2002 she has been employed in various I.T. roles and applications in manufacturing retail and finance. Her tasks have included critical analysis of documentation, data collection, statistical analysis and interpretation of information. In 1998 Gaye began a Master of Science in Computer Science at Birkbeck College part-time. She continued to work fulltime and study in the evenings. Her MSc thesis touched on her areas of interest on AI and Neural Networks. She graduated with a Master of Science in Computer Science in 2001 (Birkbeck, University of London). In 2006 she was awarded a PhD from Birkbeck College, University of London for her work on computational modelling of self control. Since then she has continued pursing her interest in Neural Networks specifically in the area of reinforcement learning and also in using computers in mathematical education.
|Mailing List: https://listserv.cs.ucy.ac.cy/mailman/listinfo/cs-colloquium|
|Sponsor: The CS Colloquium Series is supported by a generous donation from