A Learning Context Memory

Context data is often high volume, multi dimensional and dynamic; storing, reusing and learning from this information is nontrivial.

We believe there is a great challenge in being able to store signicant combinations of context, with the exclusion of redundant information. For example, a place of interest can be represented by a radius, rather than storing independently all of the GPS co-ordinates that fall within that radius.

In our opinion the requirements for context-aware systems that can adapt to behavioural changes is best achieved by combining unsupervised and reinforcement learning. A suitable denition of unsupervised learning is by [4]: ?In unsupervised learning or clustering there is no explicit teacher, and the system forms clusters or ?natural groupings? of the input patterns.

?Natural? is always dened explicitly or implicitly in the clustering system itself.? Reinforcement learning is dened by [6] as: ?Reinforcement learning addresses the question of how an autonomous agent that senses and acts in its environment can learn to choose optimal actions to achieve its goals. Each time the agent performs an action in its environment, a trainer may provide a reward or penalty to indicate the desirability of the resulting state?.

Ashbrook and Starner try to predict user movement with GPS data [1] by using a variant of the k-means clustering algorithm to cluster ?similar? GPS points. The system does not learn in real-time, which is a problem when running the system on devices with limited resources.

A number of systems exist which reason about context, they make use of a variety of techniques, for example semantic

approach, rule based [9], Markov models [8], Bayesian classiers [10], and Neural Networks [5]. Some of these techniques are able to learn in real-time which is also the case in our system.

Our system learns in real-time and stores the contextual information in the original format, but also performs data reduction. Non of the mentioned techniques full all of these properties.

In section 2 we briey introduce Articial Immune Systems; and in Section 3 we propose a learning structure based on an

AIS which is capable of forgetting potentially obsolete context, can perform data compression, and makes use of unsupervised online learning.