Reactive Systems
Yolanda Gil
CS 541, Fall 2003
(Thanks to Karen Myers from SRI International)
1
The problem with plans (I)
Attack Goliath
1.
2.
3.
4.
Gather pile of rocks
Grasp slingshot
Fire at giant
Hit on the head
2
The problem with plans (II)
Attack Goliath
1.
2.
3.
4.
Gather pile of rocks
Grasp slingshot
Fire at giant
Hit on the head
•
•
•
•
•
•
Unknown how
many stones
Unknown if stones
Unknown how
many attempts
Conditions for
termination
What if failure
3
Check state
Reactive Systems
•
•
•
•
Embedded in the real world
Have sensors and effectors
Actively test the external environment
Need to respond to events in dynamic
environments
• Failure may require aborting and generating
new response
• Do we need deliberate reasoning (planning)?
4
Outline and Informal Roadmap
• Control systems
– Networks of “variables” (arcs) and “functions” (nodes)
• Reactive Action Packages (RAPs)
– Networks of “conditions” and “tasks”
• Task Control Architecture (TCA)
– Network arranged according to “vertical capabilities”
• Procedural Reasoning System (PRS)
– Integrates planning, BDI, and reactive techniques
• Anytime algorithms
– When time is short, managing what you think about
• Other approaches and issues
5
Readings
• RAP (http://people.cs.uchicago.edu/~firby/raps)
– Firby, J “Task Networks for Controlling Continuous Processes”,
Proceedings of Artificial Intelligence Planning conference, 1994.
• TCA (http://www-2.cs.cmu.edu/afs/cs/project/TCA/release/tca.orig.html, http://www2.cs.cmu.edu/afs/cs/project/TCA/release/tca.html)
– Simmons, R. “Structured Control for Autonomous Robots”, IEEE
Transactions on Robotics and Automation, Feb 1994.
• PRS (http://www.ai.sri.com/~prs)
– Reactive reasoning and planning: an experiment with a mobile
robot, M. Georgeff and A. Lansky, in Proceedings of AAAI, 1987.
• Anytime algorithms
– Zilberstein, S. “Using Anytime Algorithms in Intelligent Systems”,
AI Magazine, 1996.
6
Control Systems: An Example (I)
Control of temperature profile for a spray deposition process. Jones, P.D.A.; Duncan,
S.R.; Rayment, T.; Grant, P.S. IEEE transactions on control systems technology
special issue on control of industrial spatially distributed processes, Sept 2003.
7
Control Systems: An Example (II)
Control of temperature profile for a spray deposition process. Jones, P.D.A.; Duncan,
S.R.; Rayment, T.; Grant, P.S. IEEE transactions on control systems technology
special issue on control of industrial spatially distributed processes, Sept 2003.
8
Beyond Stimulus-Response
• Address problems that
require a combination of:
– Coordinated activity to
accomplish tasks
– Reactivity to world
dynamics
• Situate control decisions
within an explicit,
persistent decisionmaking framework
9
Reactive Action Packages (RAP)
10
A Symbolic Discrete Task
11
Waiting for a signal to proceed
12
Concurrent
tasks
13
More
Complex
Task
Networks
14
Task Control Architecture (TCA)
• Vertical task decomposition: several taskspecific modules communicate through a
central control module
• Deliberation: top-down task-subtask,
resolve constraints
• Central control routes messages
– Inform, query, command, monitor
15
Ambler Walking Robot
16
Ambler Modules
17
Ambler Task Tree
18
TCA: Monitoring
• Central control traverses tree and handles
messages:
– asks gait planner to traverse arc,
– gait planner asks terrain mapper for elevation
map in order to take steps
– Gait planner asks leg recovery planner to place
leg, move leg, move body,
– Gait planner activates monitor whether
achieved position
19
TCA: Control
• Ordering and temporal constraints
• Delay planning constraint: goal cannot be issued
until previous task achieved
– Can do place leg planning while monitoring achieve
position
• Exception handling: error recovery modules
examine and modify task trees
– Eg: if position not achieved, add take steps subtask
20
Ambler Planning and Execution
21
An Alternative to TCA’s Vertical Capabilities:
Horizontal Layered Control
Reason about behavior of objects
Plan changes to the world
Identify objects
Monitor changes
Build maps
Explore
Wander
Avoid objects
22
Procedural Reasoning System
(PRS)
• Framework for symbolic reactive control
systems in dynamic environments
– Eg Mobile robot control
– Eg diagnosis of the Space Shuttle’s Reaction
Controls System
23
PRS: Main Features
• Pre-compiled procedural knowledge
• BDI (Belief, Desires, Intentions) foundation
• Combines deliberative and reactive features
– Plan selection, formation, execution, sensing
•
•
•
•
•
Plans dynamically and incrementally
Integrates goal-directed and event-driven behavior
Can interrupt plan execution
Meta-level reasoning
Multi-agent planning
24
PRS Architecture
User
Tasks
Procedures
Interpreter
Database
Intentions
World
25
PRS Architecture:
Database
• Contains beliefs or
facts about the
world
• Includes meta-level
information
– Eg goal G is active
User
Tasks
Procedures
Interpreter
Database
Intentions
World
26
PRS Architecture:
Tasks
• Represent desired
behavior
• Conditions over some
time interval
– eg (walk a b): set of
behaviors in which
agent walks from a to
b)
User
Tasks
Procedures
Interpreter
Database
Intentions
World
27
Expressing Tasks in a Dynamic
Environment
•
•
•
•
•
•
(! P) -- achieve P
(? P) -- test P
(# P) -- maintain P
(^ C) -- wait until C
(-> C) -- assert C
(~> C) -- retract C
28
PRS Architecture:
Intentions
• Currently active
procedures
• Procedure
currently being
executed
User
Tasks
Procedures
Interpreter
Database
Intentions
World
29
PRS Architecture:
Procedures
• Pre-compiled
procedures
• Express actions
and tests to achieve
goals or to react to
conditions
User
Tasks
Procedures
Interpreter
Database
Intentions
World
30
Representing Procedures with
Act Formalism
• Environment conditions
– Purpose (goal or condition)
– applicability criteria
• Plot
– directed graph
– partially ordered conditional &
parallel actions, loops
– Successful node execution by
achievement of node’s goals
– If no body: primitive action
Cross-Country Delivery
Cue:
(ACHIEVE (DELIVER CUSTOMER.1 GOODS.1))
(TEST
(AND
(LOCATED CUSTOMER.1 CITY.2)
(LOCATED GOODS.1 CITY.1)
(DISTANCE CITY.1 CITY.2 DISTANCE.1)
(> DISTANCE.1 1000) ) )
Setting:
(TEST
(AND
(AIR-SHIPMENT AIRCARGO.1 GOODS.1)
(LAND-SHIPMENT LANDCARGO.1 GOODS.1) ) )
Resources:
- no entry -
Metapredicates
–
–
–
–
Achieve
– Achieve-By {proc}
Test
– Conclude {effects}
Wait-Until – Use-Resource
Require-Until
(ACHIEVE
(RECORD-INVOICE
CUSTOMER.1
GOODS.1
INVOICE.1) )
Preconditions:
Propertities:
(AUTHORING-SYSTEM ACT-EDITOR)
(ACHIEVE-BY
(LOCATED
AIRCARGO.1
CITY.2)
SHIP-BY-AIR) )
(ACHIEVE-BY
(LOCATED
LANDCARGO.1
CITY.2)
SHIP-BY-RAIL) )
(ACHIEVE
(LOCAL-DELIVERY
CUSTOMER.1
GOODS.1) )
(CONCLUDE
(COMPLETED-INVOICE
INVOICE.1) )
Comment:
Long distance delivery of goods to customers
31
PRS Interpreter
Execution Cycle
1. New information
arrives that
updates facts
and/or tasks
2. Acts are triggered
by new facts or
tasks
3. A triggered Act is
intended
4. An intended Act is
selected
5. That intention is
activated
6. An action is
performed
7. New facts or
tasks are posted
8. Intentions are
updated
New Facts & Tasks
2
(overpressurized fuel-tank)
7
(ACHIEVE (position ox-valve closed))
Act Library
1
ACT2
Cue:
(TEST (overpressurized tank.1))
Act Execution
External
World
Facts
&
Tasks
6
8
5
(ACHIEVE
(position ox-valve closed))
ACT1
current
4
ACT1
Cue:
(ACHIEVE (position valve.1 closed))
Goal2
ACT8
sleeping
Goal3
ACT3
sleeping
3
Fact1
ACT2
normal
Intention Graph
32
Meta-Reasoning
• Can include meta-level procedures
– eg: choose among multiple applicable
procedures
– eg: evaluate how much more reasoning can be
done within time constraints
– eg: how to achieve a conjunction or disjunction
of goals
33
Shuttle’s RCS Malfunction Handling
RCS Controls
T
HE
P Tank P
T Temp
P Pressure
A
V alve
Talkback
C ontrol Panel
OP
B
CL
A
B
open
gpc
close
Jet
OP
R lv
V alve
R egulator
12
P
1
OP OP OP
345
RCS Jets
2
P
3
P
4
P
1
P
3
OP OP
4
5
P
Achieve:
Position valve.ox closed,
Position valve.fu closed
Cue
Test:
Alarm sounding,
RCS warning light on,
Status RCS jet.1 is failed-on,
GPC displays dir.1 for jet.1 for rcs.1
Shuttle
GPC
Achieve:
Notify "Thruster jet.1 failed-on"
Test:
High-usage of jet.1
Setting
Test:
Connected manifold.ox to jet.1,
Connected manifold.fu to jet.1,
Connects valve.fu by leg.fu
to manifold.fu,
Connects valve.ox by leg.ox
to manifold.ox,
Oxidizer-subsystem ox.1 of rcs.1,
Fuel-subsystem fu.1 of rcs.1,
Part valve.ox of ox.1,
Part valve.fu of fu.1
2
5
Jet Fail - On
Preconditions
Test:
Direction jet.1 is dir.1
345
open
gpc
close
FU
T Tank P
12
Sw itch
CL
• Automates specification and
execution of RCS malfunction
procedures.
• Reacts to changes in RCS. Ensures
safe operation while carrying out
diagnosis and remediation
procedures.
External
TASKS
Test:
Not high-usage of jet.1
MESSAGES
External
FACTS
Regulator Test
Jet Fail - On
Test:
Type jet.1 vernier
Achieve:
Notify "Thruster jet.1 failed-on
ELECTRICALLY"
Achieve:
Notify "Thruster jet.1 failed-on
INPUT CARD"
Test:
Not type jet.1 vernier
Achieve:
Pressure manifold.ox is pres.ox,
Pressure manifold.fu is pres.fu
Test:
> pres.ox 130,
> pres.fu 130
Test:
≤ pres.ox 130,
≤ pres.fu 130
Dump Propellant
TASKS
Procedure
Library
Determine new
procedures
that are eligible
for execution
Achieve:
Notify "TURN-OFF rcs.1 manifold.ox
& manifold.fu DRIVER"
Select procedures
for execution
FACTS
&
BELIEFS
Executing
procedures can post
GOALS, FACTS, &
BELIEFS
or
send MESSAGES
Jet Fail - On
Regulator Test
Multiple Tasks, Multiple Agenst
• Multithreaded operation: multiple tasks
being performed, runtime stacks where
tasks are executed, suspended, and resumed
• Supports distributed planning: several PRS
agents run asynchronously and
communicate through message passing
35
Anytime Algorithms
• Time to deliberate about events varies
• Algorithms to compute the best answers they can in the
time available
• Anytime algorithms
– Can be suspended and resumed with little overhead
– Can be terminated at any time and return some answer
– The answers returned improve with time
36
A time-dependent planning
problem
• Observe (O)
• React (E): time required to carry out reaction
of type E
• Herald (C): earliest observation time that
enables prediction of condition C requiring a
response
• Utility (C,E): utility of reacting to with E to C
• Response (C): time between having
information to predict C and C occurring 37
When Time is Short…
• Prediction time: time required to predict
event given info available
• Deliberation time: max time for committing
to a reaction (if reaction is needed)
• Reaction time: time required to react to
event
– React(E) + Response(C)
38
Deliberation
• Decision procedure D for each C: given t
time to deliberate, D returns best guess E
about how to react
• Utility(C, D(C,T))
• Deliberation scheduling:
– Given several deliberation procedures,
determine how to best allocate deliberation time
39
Utility versus time
One-shot improvement
Linear improvement,
bounded utility
Linear improvement,
unbounded utility
Diminishing returns
40
Other Approaches and Issues
• Blackboard architectures (Guardian)
• Universal plans
• Related issues covered in the course:
– Reasoning about uncertainty
– Learning
• from the environment
• Becoming increasingly reactive
41
Summary
• Control systems
– Networks of “variables” (arcs) and “functions” (nodes)
• Reactive Action Packages (RAPs)
– Networks of “conditions” and “tasks”
• Task Control Architecture (TCA)
– Network arranged according to “vertical capabilities”
• Procedural Reasoning System (PRS)
– Integrates planning, BDI, and reactive techniques
• Anytime algorithms
– When time is short, managing what you think about
• Learning and uncertainty reasoning
42