The machine learning approach that will be implemented will be discussed subsequently however it is important to note that it is a rather generic one. There is no explicit notion of dynamics, physical constraints, physical laws, behavioural structures, etc. This allows the model to be flexible and applicable to different learning environments. Sometimes, perceptual and expressive data will not fall into cognitive models that can be implemented computationally. Thus, for generalization purposes, we will not take unnecessary advantage of constraints about the domain, the tasks and the perceptual data provided to the ARL system by using hard wired cognitive models or a priori user defined models.
In addition, a probabilistic model provides a methodological framework for operating on and flexibly learning from standard data samples. It allows conditioning, sampling, computing expected behaviours and computing maximum likelihood estimates in a formal and consistent sense [5]. In addition, the underlying probabilistic mechanisms allow 'soft' associations, stochastic behaviour and continuous, unlabeled and unconstrained representations. Even though it may not be as structured or as specific as cognitive or alternative models of behaviour, it is computationally more flexible. Finally, machine learning is an automatic process with minimal effort on the behalf of the user. The desired behaviours are enacted in front of the system rather than explicitly identified by the user who must incorporate specific knowledge and skill from various domains. In other words, many variables can be estimated from data rather than manually tweaked by users.