Realtime Situational Awareness

1
Real-time Situational Awareness

• Bruce D'Ambrosio
• Oregon State University
• CleverSet, Inc.

2
Overview
OLMA

• History
• Architecture
• Algorithms
• Representation
• Multi-agent
• Learning

SSARE
ADS
RTSU
3
Pilots Associate

• Expert systems meet hard real-time
• Engineering
• Architecture
• Algorithms

• On-line maintenance agent
• Hidden state
• Fixed cycle
• Action options
• Do nothing
• Ask for addl data
• replace

4
Real-time AI Architectures

• Expert systems meet hard real-time
• Engineering
• Architecture
• Algorithms

5
Reaction and Deliberation I

• On-line
• Planning
• Problem-solving
• Off-line
• Reactive planning
• Reinforcement learning

6
Reaction and Deliberation I

• On-line
• Planning
• Problem-solving
• Off-line
• Reactive planning
• Reinforcement learning
• In qualitative belief space (Kappa reduced)

7
Reaction and Deliberation II

• Before and Beneath any deliberation is the
  continuous decision of what to do NOW (Chapman
  and Agre, Pengi)
• Temporal properties of algorithms (Horvitz)
• Anytime algorithms (Dean)
• Meta-level control (Russell)
• Optimal finite real-time agents (DAmbrosio)
• Reactive ground
• Input
• Sensors
• Deliberative state
• Actions
• Take action in world
• Initiate/modify/terminate a deliberation
• What is the optimal reactive policy given
• Finite space
• Finite deliberative computational resources.
• Previous solution is purely reactive

Quick sort
Bubble sort
8
Detour Properties of Inference Algs

• Simulation converges slowly (Pearl)
• Linear pdf entries have 2/e of mass
• In diagnostic problems (DAmbrosio, UAI93)
• In almost all problems (Druzdel, UAI)
• Estimation becomes a search problem

Incremental Term Computation
Simulation
9
Focused Partial Evaluation in OLMA

• Burgess and DAmbrosio, UAI96
• Simple fixed reactive policy
• Run deliberative n steps
• N is parameterized by problem

10
Representation I TSUGDA/SSARE

• DAmbrosio et al, SSARE, Discex II, 2000

11
Frame Ontology

• Object-oriented Probabilistic Models (PRMs)
• Normal parameter uncertainty
• New
• Existence,
• Type,
• Reference Uncertainty

12
Sample Frames - Activity
13
Representation I TSUGDA/SSARE

• DAmbrosio et al, SSARE, Discex II, 2000

14
Agent communication in RTSU

• ADS project, IET, Shozo Mori PI

15
Agent communication in RTSU

• ADS project, IET, Shozo Mori PI

16
Agent communication in RTSU
17
Real-time Situation Modeling

• Everything so far requires models.
• Where do they come from?

18
Real-time Situation Modeling
19
Summary

• Fixed propositional problem compiled off-line
  (reactive execution)
• Fixed propositional problem solved on-line
  (reactive control)
• Relational model solved on-line (reactive
  control)
• Efficient agent communication

20
Overview

• Algorithms
• Adaptive probing
• Anytime anyspace inference
• Particle filters
• Anytime replanning
• Scheduling sensor networks
• Architecture/Control
• Meta-level control
• Real-time Control Arch.
• Fine-grain control
• Real-time lookahead control
• AEDGE
• ?
• Co-evolution
• Virtual factories

• History
• Architecture
• Algorithms
• Representation
• Multi-agent
• Learning

21
Connections? Open?

• Algorithms
• Adaptive probing
• Anytime anyspace inference
• Particle filters
• Anytime replanning
• Scheduling sensor networks
• Architecture/Control
• Meta-level control
• Real-time Control Arch.
• Fine-grain control
• Real-time lookahead control
• AEDGE
• ?
• Co-evolution
• Virtual factories

• History
• Architecture
• Algorithms
• Representation
• Multi-agent
• Learning