The world s first collaborative machine-intelligence competition to overcome spectrum scarcity

Transcription

1 The world s first collaborative machine-intelligence competition to overcome spectrum scarcity Paul Tilghman Program Manager, DARPA/MTO 8/11/16 1

2 This slide intentionally left blank 2

3 This slide intentionally left blank 3

4 This slide intentionally left blank 4

5 This slide intentionally left blank 5

6 This slide intentionally left blank 6

7 This slide intentionally left blank 7

8 This slide intentionally left blank 8

9 SC2 Competition Architecture Proposers Day Overview 9

10 SC2 Competition Architecture Proposers Day Attendance for today For organizations and teams interested in participating in SC2 as a member of the competition architecture team DARPA-BAA Available in-person and online The event is not open to the press. If you are with the press, we please ask you to sign out of the Webinar and contact outreach@darpa.mil 10

11 Slides from this presentation All the slides from today s presentation are available on our website 11

12 Questions All questions must be submitted in written form Submit questions to: Those attending in person are encouraged to submit questions via Index cards provided if needed One question per submission Keep questions succinct. No multipart questions Really, one question, per or per card 12

13 Spectrum Collaboration Challenge Competitor Overview Exploit the intersection of Artificial Intelligence and Software Defined Radio To usher in a new era of spectrum sharing Based on collaborative optimization of the spectrum By designing Collaborative Intelligent Radio Networks (CIRNs) Competition Architecture Overview Developing scenarios, scoring, and competition integrity to ensure the successful execution of the worlds largest spectrum tournament 13

14 Agenda 8:00-9:00 Registration & Teaming 9:00-9:15 Orientation 9:15-9:40 Program Motivation 9:40-10:00 Colosseum (Testbed) 10:00-10:15 TA1 Scoring Methodology 10:15-10:30 TA2 Scenario Development 10:30-10:45 TA3 Integrity Development 10:45-11:10 Program Structure 11:10-11:30 Contracts 11:30-12:30 LUNCH 12:30-12:45 Answer Session 12:45-1:00 Schedule & Closing Remarks 1:00-2:00 Teaming Discussions 14

15 Motivation for the Spectrum Collaboration Challenge 15

16 The Spectrum Crunch Increasing Spectrum Demand 1000 EB/mo 4.0 EB/mo Dwindling Supply US Spectrum Auction Pricing $636M / MHz $306M / MHz $152M / MHz $23M / MHz

17 America s Cup and the beginning of frequency isolation September 1899 Marconi provides up-to-theminute reports on America s cup via spark telegraphy. October 1899 Marconi equips USN Massachusetts and New York. Interference results from simultaneous ship and shore communication. April 1900 Marconi s famous 7777 patent enables frequency tuning on transmit and receive

18 The Wireless World Today: A patchwork of planned isolation 18

19 Today Evolving access, but still isolation Evolution of spectrum access strategies 1920 s 1960 s / 1980 s 2010 s Network A Network B Network C, D, E X X X SAS Licensed Static Isolation Systems never operate at the same time, place, channel Unlicensed Blind Sharing Systems operate without information from others Dynamic Spectrum Access Coarse Dynamic Isolation Spectrum is coarsely (re)allocated Inefficient Scarce Vulnerable 20% * utilization unscalable avoidance $ / MHz Constrained and well-known operating freqs New approach needed to realize truly shared and optimized spectrum access * Executive Office of PCAST, Realizing the Full Potential of Government-Held Spectrum to Spur Economic Growth, tech. rep., Jul

20 Collaborative AI: A New Normal Vehicle-to-Vehicle Collaboration AI adapting to changing road conditions, not as a single vehicle, but as an ensemble 20

21 Tomorrow Spectrum Collaboration Only the radios themselves know their interference conditions, so they should be the ones making interference mitigation decisions x increase in efficiency Autonomous collaboration for efficient spectrum sharing 21

22 Radio Network Autonomy Requires Learning Autonomy: radio network self-optimizes to changing operating conditions Autonomy requires a radio network to repeatedly decide its: frequency waveform antenna direction, hop pattern, time slot, Each decision depends in complex ways on the current situation number and location of devices propagation channel between each device pair other network types and operating characteristics current spectrum use by other networks Complex dependencies create combinatorial decision space growth Hard-coded algorithms can only achieve limited autonomy Broaden scope of autonomy through learning 22

23 Learning and Collaboration Are Synergistic Learning benefits from information sharing The environment includes other radio networks Accurate information about their needs, outcomes, and plans reduces the combinatorial space If interference occurs, accurate information about why it happened accelerates discovery of successful strategies Collaboration benefits from learning Networks seek to optimize others performance in addition to their own Much faster combinatorial explosion Hard-coded algorithms are even less effective for collaboration than for single-network autonomy Full state-space isn t known at design time 23

24 Spectrum Collaboration Challenge Summary An open competition To develop radio networks Which can thrive in the spectrum without allocations Which learn how to adapt across multiple degrees of freedom Which collaboratively optimize the total spectrum capacity, moment-to-moment 24

25 What does a team have to do? Each team develops a Collaborative Intelligent Radio Network (CIRN) AI & Machine Learning NET MAC PHY Firmware Software Standard Standard Radio Standard Radio Node Radio Node Node Learning Control Algorithms Complete Radio Stack Instantiated in SW/FW Using the SRN hardware Winning design is the one that best shares spectrum with any network(s) in any environment without prior knowledge CIRNs are evaluated in a collaborative, tournament-style competition 25

26 SC2 Competition Structure SC2 Competition events are round-robin tournament style Each match will pair up to 5 competitor teams together (i.e., ensemble) Teams must demonstrate the ability to work with numerous ensembles Overall score determined as a composite across the entire tournament Ensuring Fairness of the Competition Ability to create identical RF conditions across matches Play in ensembles with all other eligible radio networks Scrimmaging to prepare for tournaments Monthly scrimmages will be established within small divisions 26

27 Match Overview One node per network serves as a gateway Collaboration takes place over internet-like infrastructure connected to the gateway (models realistic internets) IP Traffic Team 1 Team 2 Team 3 Team 4 Team 5 Ensemble of up to 5 teams placed in arena Incumbent Each node is given IP traffic Sources and destinations are contained in the same network Traffic will emulate multiple canonical types Arena may also contain other Non-Collaborative Radios (NCR): Incumbents Jammers Radio environment emulated in real-time: Large-scale path loss Multipath & Doppler Channel correlation Motion 27

28 SC2 Match Detail Establish initial operation condition Respond to change 0 min 5 min Note: Times shown are exemplary. Match durations and change event frequency may be different than shown. Change Event 5-30 min Movement Available bandwidth Priority Traffic characteristics Non-collaborative radios Collaboration channel QOS Propagation environment Ensemble composition 28

29 What does a match look like T 0 Team 1 Team 3 Team 2 Team 4 Team 5 Teams work together to conquer obstacles and find effective communication strategies 29

30 What does a match look like T n 30

31 What does a match look like T final Just one example of obstacles that may be presented during a match 31

32 Scoring Match Score = ff(performance score, collaboration score) A function of: Goodput Prioritization Fairness Quality of Service A function of: Other networks performance scores Interference to incumbents Tournament Score = weighted sum of match scores Scoring formula & mechanisms are under development Specifications will be finalized approximately 2 months prior to competition event 32

33 Collaboration Protocol Information exchange between networks Important byproduct of SC2 No such protocol exists today What is it? An extensible protocol Large majority of fields are optional Who is designing it? Everybody Living, growing protocol, to experimentally add to DARPA will periodically merge all the versions in the wild to create a new baseline 33

34 5 Elements of a Collaborative Intelligent Radio Network Reconfigurable radio Adaptability along dimensions of time, frequency, space, code, waveform, MAC scheme, network, etc. Intelligent radio characteristics Understanding: Understand and characterize signals to infer the conditions of the local RF environment through noisy observations Collaborate Reasoning: Reason about how to take sequential actions to result in successful communication, taking into account the effect the action may have on others using the same spectrum Contextualization: Contextualize what the system has already learned to be able to rapidly overcome changes and new challenges by leveraging and transferring previous knowledge to these new problems Collaboration: Learn how to collaborate with previously unknown radio systems by discovering what information is valuable to them to optimize the overall joint outcome while minimizing the cost of collaboration 34

35 SC2 Competitor Summary DARPA seeking innovative approaches for a new breed of radio network Collaborative Intelligent Practical Incremental evolution of existing approaches are not sought Open loop only (e.g., non-collaborative) approaches Whitespace hunting Approaches which rely on hardware other than the SRNs provided Paradigm shifting approaches do not have to be constrained by traditional comms architectures (military and commercial) 35