Information Technology Reference
In-Depth Information
Online performance.
The reasoning algorithms for activity recognition and situation
detection must be able to perform online analysis. This means that a trade-off must be
made between the model complexity and the computational complexity of the
algorithm.
Modelling of a-priori knowledge vs. learning.
The layout and equipment of
environments in which the system will be deployed can vary largely, as can the user's
performance of activities. This makes it impracticable for the reasoning to rely solely
on models of a-priori knowledge. To adapt the models over time to the individual user
the different types of machine learning have to be analysed and integrated. Some of the
algorithm types are already used for AAL applications:
Supervised learning
: Generates a function that maps inputs to desired outputs. For
example, in a classification problem, the learner approximates a function mapping a
vector into classes by looking at input-output examples of the function. Reasoning
models can be adapted to the characteristics of the specific environment and the
specific user during an explicit training phase
Unsupervised learning
: Models a set of inputs, like clustering and is for example used
to analyse the reaction of person to medication.
Semi-supervised learning
: Combines both labeled and unlabeled examples to generate
an appropriate function or classifier.
Reinforcement learning
: Learns how to act given an observation of the world. Every
action has some impact in the environment, and the environment provides feedback in
the form of rewards that guides the learning algorithm. e.g.: reinforcement learning is
used to adapt schedule management supporting elderly people in planning and
executing activities.
Transduction
: Tries to predict new outputs based on training inputs, training outputs,
and test inputs.
Learning to learn
: Learns its own inductive bias based on previous experience.
Pareto-based multi-objective learning
: a Pareto-based approach to learning that
results in a set of learning models, which typically trade off between performance and
complexity.
On the other hand, it is necessary to provide a-priori knowledge to the reasoning
system, e.g. the medical knowledge modelled by domain experts. The reasoning system
should be able to handle this trade-off between user adaptability by learning and
explicit modelling of a-priori expert knowledge.
Reasoning approaches can be distinguished according to two major aspects,
namely the underlying knowledge representation formalism and the type of semantics
used. The underlying formalism largely influences the expressive power: approaches
based on first-order logic are much more powerful than propositional formalisms,
whereas extensions of description logics or relational schemes lie between these
extremes. On the other hand, semantics determines how information (i.e. valid
sentences in a knowledge representation language) is interpreted. Whereas there are
countless syntactic variations in knowledge representation systems, most existing
formalisms can be seen to follow either a proof-based or a model-based semantics. In
the following, the different approaches are described briefly.
Rule-based reasoning
gives no inherent support for reasoning of incomplete data or
the handling of uncertain information (probabilistic information). While extensions
