Site Report: USRA Quantum Computing AI Lab · 2019-04-10 · Site Report: USRA Quantum Computing AI...
Transcript of Site Report: USRA Quantum Computing AI Lab · 2019-04-10 · Site Report: USRA Quantum Computing AI...
Davide Venturelli – QUBITS EUROPE – March 26th 2019
Site Report: USRA Quantum Computing AI Lab(NASA-Google-USRA collaboration)
Davide Venturelli,Quantum Artificial Intelligence Laboratory,USRA:RIACS Science Operations Manager
and Quantum Computing Task LeadResearch Scientist @ NASA Ames Research Center [email protected] ([email protected])
Fund
ing by
NASAEleanor RieffelJeremy FrankChris TeubertRupak Biswas
USRADavide VenturelliZhihui WangStuart HadfieldFilip WudarskiEugeniu PlamadealaJeffrey MarshallNorman TubmanVanesa Gomez-GonzalezDavid Bell
Fellowship/VisitingBryan O'Gorman(Tad Hogg)(Boris Altshuler)
https://ti.arc.nasa.gov/tech/dash/groups/physics/quail/
SGTSalvatore MandràGianni MossiWalter VinciAdam Max Wilson
External Collaborators (2018)Kyle Jamieson, Minsung Kim (Princeton), Alexei Kondryatev (Standard Chartered Bank), Stefano Casalegno (FFSS), Bibek Pokharel (USC), Hemant Shukla (Siemens)
Recent Reference papers: The power of pausing: advancing understanding of thermalization in experimental quantum annealers (J.Marshall, D.Venturelli, I.Hen, E.G.Rieffel) – arXiv:1810.05881 (2019) Reverse quantum annealing approach to portfolio optimization problems (D.Venturelli, A.Kondratyev) – Quantum Machine Intelligence Journal, arXiv:1810.08584: (2019) Leveraging Quantum Annealing for Large MIMO Processing in Cloud-Based Radio Access Networks (M.Kim, D.Venturelli, K.Jamieson) – to appear Performance of Quantum Annealers on Hard Scheduling Problems (B. Pokharel, D. Venturelli, E.G. Rieffel) – to appear
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Outline
Brief history of the use of the D-WAVE 2000Q at NASA Ames Phenomenological model of annealing dynamics Experiments with the Pause Performance Distribution
Embedded Spin Glasses Forward annealing versus Reverse Annealing Performance examples on embedded problems Wireless Networks Maximum Likelihood Decoding Financial Portfolio Optimization
Conclusions and Future Directions
1
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Brief History of the D-Wave Quantum Annealers
ILIAC IV, NASA Ames Research CenterFirst massively parallel computer 64, 64-bit FPUs and a single CPU 50 MFLOP peak, fastest computer at
the time
Finding good problems and algorithms was challenging.«Would computers ever be able to compete with wind tunnels?»
NASA, Google and USRA formed a three-way collaboration focused on Artificial Intelligence and Quantum Computing (2012-present)
Universities,Industry,Startups, NSF
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Update on Usage of the NASA Ames D-Wave Machine
Quantum RFP
Competitive SelectionsCycle 1 (512 qubit processor): 8 of 14 selected – 57%
Cycle 2 (1152 qubit processor): 10 of 15 selected – 67%Cycle 3 (2048 qubit processor): 15 of 19 selected – 79%
Diversity of Selected OrganizationsApprox 60% Universities + 40% Industrial Research Organizations Approx 60% U.S. Organizations + 40% International OrganizationsComputer Science, Physics, Mathematics, Electrical Engineering, Operations Research, Chemistry, Aerospace Engineering, Finance
Diversity of ResearchQuantum Physics -> Algorithms -> Applications
Machine Learning for Image Analysis, Communications, Materials Science, Biology, Finance
RFP CYCLE 1 & 2 SELECTIONS
RFP CYCLE 3 (extract)
https://tinyurl.com/USRA-RFP2019
Davide Venturelli – QUBITS EUROPE – March 26th 2019
D-Wave Two™ D-Wave 2X™ D-Wave 2000Q™512 (8x8x8) qubits “Vesuvius” 1152 (8x12x12) qubit “Washington” 2048 (8x16x16) qubit “Whistler”
509 qubits working – 95% yield 1097 qubits working – 95% yield 2038 qubits working – 97% yield
1472 J programmable couplers 3360 J programmable couplers 6016 J programmable couplers
20 mK max operating temperature (18 mK nominal)
15 mK Max operating temperature (13 mKnominal)
15 mK Max operating temperature (nominal to be measured)
5% and 3.5% precision level for h and J 3.5% and 2% precision level for h and J To be measured.
Annealing time 20 µs Annealing time improved 4x (5µs)Readout time improved (120µs)
Annealing time improved 5x (1µs)Initial programming time improved 20% (9 ms). Extented J, anneal offset, pause and quench features. (+h field schedules [2019])
■ ■ The D-Wave machine at NASA Ames 5m
Davide Venturelli – QUBITS EUROPE – March 26th 2019
D-Wave Two™ D-Wave 2X™ D-Wave 2000Q™512 (8x8x8) qubits “Vesuvius” 1152 (8x12x12) qubit “Washington” 2048 (8x16x16) qubit “Whistler”
509 qubits working – 95% yield 1097 qubits working – 95% yield 2038 qubits working – 97% yield
1472 J programmable couplers 3360 J programmable couplers 6016 J programmable couplers
20 mK max operating temperature (18 mK nominal)
15 mK Max operating temperature (13 mKnominal)
15 mK Max operating temperature (nominal to be measured)
5% and 3.5% precision level for h and J 3.5% and 2% precision level for h and J To be measured.
Annealing time 20 µs Annealing time improved 4x (5µs)Readout time improved (120µs)
Annealing time improved 5x (1µs)Initial programming time improved 20% (9 ms). Extented J, anneal offset, pause and quench features. (+h field schedules [2019])
■ ■ The D-Wave machine at NASA Ames 5m
B.Bokharel, D.Venturelli, E.Rieffel (2019)
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Recap of D-Wave 2000Q parameters, schedules
Pause
Coherent Evolution does not describes the dynamics:
+ coupling to a bath(s) + dynamics of the bath(s)
Davide Venturelli – QUBITS EUROPE – March 26th 2019
J.Marshall, DV, I.Hen, E.Rieffel (2019)
Quasi-static evolution
■ ■ Phenomenological model of annealing dynamics
Initial evolution locked in the ground state
At Equilibrium with the environment
Out of equilibrium with the environment(trying to cool down)
Frozen, transitions suppressed. State is some NEQ distribution
Equilibrium Excess Energy <E>-E0
10m
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Experiments with the pause: pause location
Out of equilibrium with the environment(trying to cool down)
Theory advances on open system quantum annealing: Weak Coupling Limit
o Albash, Boixo, Lidar, Zanardi, NJP 2012 Solvable Toy Models (nonperturbative)
o Smelyanskiy , Venturelli et al., PRL 2017o Kechedzhi , Smelyanskiy, PRX 2016
Realistic Simulationso Boixo , Smelyanskiy et al, Nature 2016o Smirnov, Amin , NJP 2018
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Experiments with the Pause: pause duration
783 qubits10k num reads1ms anneal time
12 qubits10k num reads1ms anneal time
50 instances, planted problems10k num reads1ms anneal time, 100ms pause
The pause always improves results(order of magnitudes TTS improvement!)
15m
Logarithmic dependence
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Experiments with the Pause: Distribution
Box plot of the difference between the pause point for which the maximal R2
value is found (i.e., the closest fit to a Boltzmann distribution based on G=0–estimated by entropic sampling technique), and the optimal pause point, with problem size.
55 instances, planted 10k num reads1ms anneal time100ms pause time
Classical Boltzmann Distribution
Best fit to a quantum Boltzmann distribution
12 qubits10k num reads1ms anneal time1000ms pause time
20m
Conjecture→at optimal pause point, the distribution is always well approximated by a classical Boltzmann for large N.
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Take home messages from pausing anneal in native sparse Ising problems
1. We experimentally observed orders of magnitude improvement in performance with respect to unpausedannealing for certain problems of the planted-solution type by the use of a pause at the optimal location.
2. The optimal pause location is found to occur after the location of minimum gap, as expected; we conjecture pausing after the minimum gap allows for the ground state to repopulate after dissipative transitions which occur during the region of the minimum gap.
3. we provide evidence suggesting thermalization to a classical Boltzmann distribution is occurring in problems containing up to 500 qubits.The temperature of this classical distribution however is different from the physical temperature of the device.
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Pause and Reverse Annealing Protocol
One more «knob» to try in empirical research heuristics.o D-Wave Whitepaper 14-1018A-A (2018)o Ottaviani et al. arXiv:1808.08721v1 (2018)
Some theoretical research on RQA:o Ohkuwa et al. PRA (2018)o Passarelli et al. arXiv:1902.06788 (2019)
• 10-100x improvement• Circumvent some bottlenecks if start
is good enough
Which initialization? Where to stop? For how long?
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Benchmarks on Applications Embedded on Fully-Connected Graphs
MIMO Wireless Maximum-Likelihood Decoding
Fund-of-Funds Portfolio Optimization
Decoded signal all possible signals: strings of symbols
Received signal
Wireless channel estimated via preambles – changes at ms scale, noisy
Asset preference (sharpe ratio bucket) Asset correlation matrix
Boothby et al.(2015)
Up to ≈40 users embeddable for Quadrature-phase shift keying (QPSK) [4 symbols alphabet]
+adding noise
Up to ≈60 assets embeddable for a fully-correlated problem.
Median Time to solution at 99% confidence
TTS = log(0.01)/log(1-P)
30m
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Wireless Decoding Problem: Embedding Parameter Setting, Anneal Time Optimization, Pause Location
In the paper we report BPSK, QPSK, 16QAM and with AWGN –here only QPSK.
Tests done on 10 random-H, random-y instances, 50k anneals minimum.
First time applied problem use the «extended J» feature
Best anneal time 1ms
Majority voting to decode chains representing logical variables
We look for the best pause for all parameter settings – one order of magnitude in TTS can be gained. (Only small pauses are relevant for the C-RAN MIMO Wireless problem)
( 18 users )
The expected best BER after Na anneals:
TTB10-6
MIMO Wireless Maximum-Likelihood Decoding
M.Kim, D.Venturelli, K.Jamieson(2019)
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Portfolio Optimization Problem: Greedy Pre-processing
Tested 30 instances – random parameters from NAV data statistics
Run only the instances that don’t solve
Davide Venturelli – QUBITS EUROPE – March 26th 2019
Results consistent with Venturelli et al. PRX 2015, Hamerly et a. 2019
Reverse Anneal doesn’t change the setting.
■ ■ Portfolio Optimization: Embedding Parameter Setting, Reverse Pause Optimization
JF and pause location can be set independently Shorter pause is more performant (but non-zero)
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ ResultsFund-of-Funds Portfolio Optimization
Venturelli, Kondryatev Quantum Machine Intelligence Journal 2019 (to appear)
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Other NASA-Sponsored Project : Convergent Aeronautics Solution
Feasibility study: Using quantum-classical hybrids to build a secure jam-free network to ensure the availability of a UAS Traffic Management (UTM) network against communication disruptions
Kopardekar, P., Rios, J., et. al., Unmanned Aircraft System Traffic Management (UTM) Concept of Operations, 2016
Future • Higher vehicle density• Heterogeneous air vehicles• Mixed equipage• Greater autonomy• More vulnerability to communications disruptions
Explore quantum approaches to• Robust network design• Track and locate of a moving jammer • Secure communication of codes supporting anti-jamming protocols
Harness the power of quantum computing and communication to address the cybersecurity challenge of availabilityUSRA quantum physicists are key team members
• Joint with NASA Glenn, who are working on Quantum Key Distribution (QKD) for spread spectrum codes.
Davide Venturelli – QUBITS EUROPE – March 26th 2019
■ ■ Conclusions
Open Positions: Senior Scientist (10+ yrs experience) Scientist (3+ years postdoc experience) Associate Scientist (postdoc level) Grad Student Intern (3-8 months)
Open Call to use the D-Wave 2000Q (free)https://tinyurl.com/USRA-RFP2019
Submit to Quantum Machine Intelligence Journal (Springer)https://www.springer.com/engineering/computational+intelligence+and+complexity/journal/42484
[email protected]([email protected])
(2017 – 3 spots)Stuart Hadfield (Columbia)David Roberts (MIT)Bibek Pokharel (USC)
(2018 – 6 spots)Benjamin Villalonga-Correa (UIUC)Aniruddha Sabat (U Maryland)Sasha Nanda (Caltech)Jeffrey Marshall (USC)Riccardo Mengoni (Univ. Verona)Zhenyi Qi (U Michigan)
(2019 – 8 spots)Andrea di Gioacchino (Univ. Milano)David Bernal (CMU)Minsung Kim (Princeton). . .
Interest: use and modeling of advanced features of the D-Wave machine.