Upload
justin-ray
View
216
Download
0
Tags:
Embed Size (px)
Citation preview
Controlling and Configuring Large UAV Teams
Paul Scerri, Yang Xu, Jumpol Polvichai, Katia Sycara and Mike Lewis
Carnegie Mellon University and University of Pittsburgh
Context• Aim to build large heterogeneous teams for
complex tasks– Robots, agents, people– 10,000s of actors
• Multiagent version of Belief Desire Intention approach to autonomous behavior– Builds on key abstraction of a Team Oriented Plan
• Defines the activities that must take place and interactions between those activities
– Supported by extensive theoretical (logical) work• Key algorithms are NP-complete or worse
– Heuristics required for scalability
WWWAgent ProxiesFor People
InformationAgents
Ontology-basedMatchmakers
Specific Target Application: Wide Area Search Munitions (WASMs)
• Part munition, part unmanned aerial vehicle– Single use– Variety of sensors– Limited fuel supply, approximately 30 minutes– Communicate with each other, manned aircraft
• Concept of Operations (under development)– Small number of manned aircraft – Potentially other ground forces– 100s of WASMs, performing a variety of missions
• Attack• Search• Battle damage assessment • Decoys• Communication relays
• Flight test planned September, 2005– 1 real and 3 simulated WASMs
Token-Based Coordination• Token: self contained packet capable of being sent around
team– Information content– Control content
• Local models– Team members use receipt of tokens to create local models of
other team members• What sorts of things are they/are they not working on?• What sorts of things might they need to know?
– Local models are used to improve the routing of future tokens
• Token “flows” implement coordination• No brittle, non-scalable “message protocols”
Token Based Algorithms
• Plan instantiation
• Removal of duplicate plans
• Role allocation
• Information sharing
• Resource allocation
• Sensor fusion
• Recovering from faulty sensor readings
Token Based Algorithms• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
Information about environment passed around in tokens, when agent receives tokens matching plan pre-conditions plan is initiated
Very high probability all applicable plans are initiated
Liao et al, 2004
0 200 400 600 800 10000
1000
2000
3000
4000
5000
6000
7000
8000
Number of Team Members
Con
flict
Res
olut
ion
Mes
sage
s
AlwaysLocalRandom
Token Based AlgorithmsInformation about initiated plans shared in tokens
Very high probability some agent gets to find out about any duplicate plans
Liao et al, 2004
• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
5101520
010
2030
0
0.2
0.4
0.6
0.8
1
Size of Subteam ASize of Subteam B
Pro
babi
lity
Token Based AlgorithmsResponsibility to perform a role encapsulated in token
Only team member with token can perform role
Team member must have capability > threshold to perform role
Threshold calculated from estimates of likely role allocation outcome
Okamoto, 2004
• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
1
10
100
1000
10000
100000
1000000
10000000
10 120
230
340
450
560
670
780
890
1000
1110
1220
1330
problem size (#agents = #roles)
com
mu
nic
atio
n p
er r
ole
(m
essa
ges
)
DSA
LA-DCOP, 0.0 threshold
LA-DCOP, 0.5 threshold
Greedy
Token Based AlgorithmsLocally sensed information shared via token to get information to team members performing role effected by information
Does not require sensing agent to know who needs information
Xu et al, 2004
• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
0
10
20
30
40
50
60
70
80
90
Mes
sage
s
Random Small World Grid Scale Free
Network type
Without efficiency enhanced Algorith With efficiency enhanced Algorith
Token Based AlgorithmsAccess to resource represented by token
Only team member with token can use resource
Team member must have need > threshold to keep resource
Threshold changes dynamically as it moves around the team, seeing resource need
• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
Token Based AlgorithmsUncertain sensor readings encapsulated in tokens and sent around the team
Very high probability that at least one team member gets related sensor readings and fuses for higher confidence
• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
Token Based AlgorithmsAssumptions used to make decisions are attached to tokens resulting from that decision
Very high probability agents with contradictory information see the assumptions and initiate checking process
• Plan instantiation• Removal of duplicate
plans• Role allocation• Information sharing• Resource allocation• Sensor fusion• Recovering from
faulty sensor readings
Layered Team Member Architecture• Bottom: Local Reasoning Team
member’s local actions are restricted by the tokens they have– E.g., without resource token, cannot use
resource• Middle: Coordination Reasoning
Movement of tokens around the team implements the coordination– Flows of tokens– Local models inform token routing
• Top: Meta-reasoning Ensures token flows work effectively
Token
Local
RoutingModel
Meta
Synergies Between Token Algorithms
• Overall performance depends on how well tokens are routed– Team members use local models to improve routing
• Observation: Execution of algorithms shares information that other algorithms can exploit– E.g., if role for strike near Pittsburgh was allocated to WASM X,
then air space resources around Pittsburgh likely of interest to WASM X
• Implementation: Use all tokens to improve local models– E.g., role tokens change local models, resource tokens move
according to local models
• Result: When all algorithms are working together the system actually performs better
Implementing Synergies - Example
• When receiving an role token from neighbor i– Probability of sending information token to i
changed proportionally to similarity between “plan initiated” token and information token
– Probability of sending resource token to i changed inverse proportionally to similarity between role and resource token
Meta-Reasoning for Token Flows
• Specific tokens are meta-reasoned about– Identify tokens that are not behaving as expected
– Bring to human attention
– Examples• Tokens that “live” too long
• Tokens that travel too much
– See Scerri et al, AIAA 2004
• Overall flows can be controlled to optimize for specific criteria– By controlling the flows, we control how the
coordination works
Neural Networks for Modeling and Controlling Token Flows
• Use a simple input/output feed-forward neural networks to represent a team performance model
– Three-layer FFNN is capable of representing any arbitrary functions
• Extend to Dynamic Networks concept to cope with dynamic behaviors– This kind of network enlarges the capacity to deal with non-
deterministic problem
• Learn the model with genetic algorithms methodology– Excellent for dealing with very huge and noisy training data set
– Based on around 1,000,000 simulation runs
Configuring Algorithms with NN• Offline
– Change environment and algorithm parameters, observe expected performance
• Online– Use NN “in reverse” to find parameter settings
for specific optimization criteria• As environment changes• As requirements change
– Allows tradeoff between performance of all algorithms
Configuration Interface
Results: Token Based Coordination
0
50
100
150
200
250
300
350
400
450
500
0 100 200 300 400
Rounds
Rew
ards
(The
mor
e the
bet
ter)
Ranom Algorithm Plan Initiation Role AllocationResource Sharing Integrated Coordination
15000
20000
25000
30000
35000
40000
45000
50000
55000
60000
150 200 250 300 350 400 450
RewardsM
essa
ges
(The
less
the b
ette
r)
Random Plan Initiation Role Allocation
Resource Planning Integrated Coodination
Total Reward Messages
Fully Integrated Random
Results: Configuring Algorithms
Results: Online Control
Machinetta: Bringing it all Together• Encapsulate token-based coordination approach into
reusable software module– Called a proxy
• Proxies provide domain independent coordination algorithms– Do not provide domain specific communication and interface code– Available in the public domain
• Machinetta used to demonstrate coordination of up to 500 distributed, heterogeneous team members in several distinct domains– Demonstrates that token based coordination is feasible – Biggest teams developed to date?
Conclusions and Future Work• Token-based coordination as a feasible
alternative paradigm for large teams• Additional layer over token flows gives
high levels of control
• Future Work– Can we make more precise mathematical
models?• Markov chains?