Motivation: Composite human/agent systems Time-aware

Co-ordinating Heterogeneous Interactions
in Systems Composed of
Active Human and Agent Societies
Konstantinos Prouskas and Jeremy Pitt
ESAW 2002, Madrid, Spain
Intelligent and Interactive Systems Group
Imperial College of Science, Technology and Medicine
Overview
 Motivation: Composite human/agent systems
 Time-aware agent architecture
 Time-Aware Layer
• Availability & Availability Functions
• Interaction Scheduling
o Availability Function Representation
o Selective Sampling Algorithm
 Evaluation
 Conclusions & future work
Motivation
 Composite human/agent systems
 Humans and agents as equal participants
• Open environment
• Situated System
• Heterogeneous interactions
 Co-ordinating agent-agent and human-agent interactions
Time-Aware Agent Systems
 A real-time architecture for time-aware agents
• Handles soft, hard & non-RT interactions
• Makes transformations between ‘human’ and ‘agent’ time
• Unifies & co-ordinates human-agent and agent-agent interactions
 Realisation in the agent process interaction language
(APRIL)
Time-Aware Layer
 Party Handles address human and agent parties uniformly
 Task Representation hierarchically decomposed sequences of either
processing and interacting
 Availability Functions express a party’s degree of presence within the
system
 Unified Time Representation transforms disparate time
representations, scales and calendars into a common representation
 Interaction Scheduling predicts and guarantees initiation and
completion times of interactions
Availability & Availability Functions
 The availability of a party is its degree of ‘presence’ in the system, as
far as interactions with other parties are concerned
• Determined by the periods during which parties are reachable via the
interconnection network
 Mathematically expressed as a probability
 Availability Functions function as look-up tables that provide a
statistical measure of availability at any point in time
Interaction Scheduling
 Availability determines interaction initiation and
completion times
• System uses availability for interaction scheduling
• Co-ordinate interactions such that completion time is known or
can be statistically predicted or guaranteed in advance
 On-line vs. Off-line processing
 Interaction Scheduling involves searching through
availability function values
Scheduling Interactions
 Availability of party j at time t, denoted Vj(t)
 The problem (P1): given
•
•
•
•
A function f(t)
A reference point tref
A condition C on f
Find earliest tx after tref that satisfies C
 Looking for Vj(tx) > g
• Where g is guarantee level that an interaction can be
initiated or completed
 Sampling has several problems related to
• Step size
• Iterative Process
Availability Function Representation




Stylised, analytic representation (rather than numeric)
Hierarchical operator trees
Four primitive nodes: Span, Bound, Repeat, Offset
Compositional node: addition, modulation, selection etc.
Selective Sampling Algorithm
 Minimises number of points that need to be sampled when
searching through availability functions
 Relevant points: Points at which there is a potential
change in the value of availability
• Two stages generate relevant points
 Takes advantage of piecewise constancy and repetition
 If a solution (to P1) exists, it will be found
Evaluation
 Real-time performance
• Availability Function representation: Bounded amount of space,
O(1) access to availability
• Selective Sampling Algorithm: O(n2) complexity
 Practical application results
• Workflow management system
Summary & Future Work
 Summary
• Time-Aware Layer co-ordinates (schedules) agent-to-agent and
human-to-agent interactions
• Interactions are unified by abstracting away from their temporal
representation, temporal scale and class of parties they involve
(humans or agents)
• Availability drives interaction scheduling
• Availability Function representation and Selective Sampling
Algorithm give real-time performance
 Future work
• Large-scale distribution
• Interplay between computational and interactional scheduling
• Validation of Availability Function generation
 Acknowledgements
• EPSRC ROPA Award Tara2 (GR/24940)