A Neural Network Model of Causative Actions

Motor Learning;Neural Networks;Language;Reinforcement Learning Rewards

Lee-Hand, Jeremy Sein Ong ; Knott, Alistair

University of Otago

关键词: Motor Learning;    Neural Networks;    Language;    Reinforcement Learning Rewards;   

Others  :  https://ourarchive.otago.ac.nz/bitstream/10523/4549/6/LeeHandJeremy2013MSc.pdf

美国|英语

来源: Otago University Research Archive

【 摘 要 】

Many of the actions we perform are defined by the effects they bring about, rather than as stereotypical sequences of motor movements. For instance, to open a door we must perform an action which results in the door opening. This thesis is a study of causative actions of this kind. I first introduce the class of causative actions, reviewing evidence for their existence from psychology, neuroscience and linguistics. I then present a computational model of motor control which can learn how to perform causative actions. The model I propose is an extension to an existing model in the literature (Oztop et al., 2004). In Oztop;;s model simple reach-to-grasp actions are learned through reinforcement using touch sensations which are considered to be intrinsically rewarding. In my extension, I propose that observed external events can also function as rewards if they are observed while the agent is executing a motor action and attending to the object being acted upon. I demonstrate the feasibility of this proposal in an implemented neural network model of causative actions. The model is also novel in that it does not require the trajectory of the hand to be precomputed. I conclude by discussing possible links between my model of causative actions and an account of the syntax of causative constructions in human language.