Autonomous "memory based learning":
Agent observes user actions
For every new user 'situation':
Salient features of application data & interaction
are weighted (0-1)
Features compared with library of previous experience
Agent chooses a new action based on 'do-it' and 'tell-me' thresholds
Action either: automation, suggestion or do nothing
Algorithm overview:
PRIOR TO OPERATION:
User sets tell-me & do-it thresholds
Library search scope set
INPUT:
A new user event occurs (perhaps a new e-mail has arrived)
OUTPUT:
Select action A via learning techniques
Assign a confidence value C to that action
if C > do-it threshold then
Execute action A
Add to list of automated actions that can be inspected by user
If user indicates A was incorrect, ask user to adjust contributing
weightings
else if C > tell-me threshold then
Suggest action A
else
consult other agents for help and repeat with
A' and C'
else do nothing