intelligent user interfaces

boorishadamantΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 4 χρόνια και 9 μέρες)

43 εμφανίσεις

From context sensitivity to
intelligent user interfaces

Requirements for learning agents

Jarmo Korhonen

8.10.2002

Overview


Machine learning


Software agents


Role of agents


Implementation requirements


Sensors


Actions


Use of learning results

The Incredible Learning Machine


Tasks:


Classification,clustering


Prediction


Modeling


Algorithms


Neural networks


Genetic algorithms


Bayesian learning


etc.

Definition:

The ability of a device to

improve its performance

based on its past

performance

Software Agents

For user, Software Agent is:

An artificial agent which
operates in a software
environment.

One that is authorized to act
for another. Agents
possess the
characteristics of
delegacy
,
competency
,
and
amenability
.

In AI tech., Software Agent is:

"An agent is anything that can
be viewed as perceiving its
environment through
sensors and acting upon
that environment through
effectors." Russell & Norvig

Basically, agent has sensor,
actors and goals.

Problems with ML in HCI with SA


ML needs to process all instances at once


ML requires large amounts of data


ML requires suitable amount of features


ML assumes static feature space


User input difficult to apply to ML


ML requires clear goals


Mistakes need to be corrected by expert

Machine learning in UI


Must learn quickly


two to five samples


Continuous environment


must decide
what is a sample from huge feature space


Incremental and sequential


order is
important


Sustainable


incremental learning


Reversible, ability to forget

Role of Agents


Taking initiative


Visibility


what is the
agent doing


Synchronizing with user


Trust


required for delegating

Sensors


Context, intent, emotion etc.: all are
indirect sensors


Direct sensors: user actions,
software/device internal state


There must be a mapping between direct
sensors and needed indirect sensors


Learning can be done with either


but feature space for direct sensors is huge

Actions


Agent has a set of possible actions


Agent has a goal


Select action that go towards the goal


In user interface agents, actions may be


Suggestions to user


Anticipate the actions of user


Operations on the behalf of the user

Results of learning


The learning should be used for something


Change the user interface


context
-
sensitivity, adapting to different users


Agent role is assistant


Automating tasks


Repetitive tasks, tasks with long duration


Agent role is autonomous

Conclusions


Learning technology needs to be improved


Take hints from user


Constrain automatically the feature domain


Learn incrementally and sequentally


Agents still need to be tailored to the task