Recommended
More Related Content
What's hot
What's hot (19)
Similar to Dario izzo - Machine Learning methods and space engineering
Similar to Dario izzo - Machine Learning methods and space engineering (20)
More from Advanced-Concepts-Team
More from Advanced-Concepts-Team (20)
Recently uploaded
Recently uploaded (20)
Dario izzo - Machine Learning methods and space engineering
- 1. Machine Learning methods and Space Engineering By Dario Izzo
- 2. Mission Created in 2002 “to monitor, perform and foster research on advanced space systems, innovative concepts and working methods” 2
- 3. too immature for regular ESA programmes or projects concepts, techniques & scientific domains with no/weak links to space emerging from cutting-edge basic scientific research topics on which ESA is expected to have a position biomimetic approaches to engineering, brain-machine interfaces, liquid breathing, curiosity cloning, peer-to-peer computing, crowd sourcing gaming, innovation diffusion and dynamics mathematical global optimisation techniques, cloud-based uncertainty modelling, helicon thrusters, pure general relativistic approach to GNSS constellation design, vibrating systems in general relativity, metamaterials in the optical frequency range, distributed/swarm intelligence, laser filamentation planetary protection research, space nuclear power sources, asteroid deflection, liquid ventilation, pulsar navigation, biomimetic drilling solar power from space, torpor/hibernation, asteroid deflection, active removal of space debris, novel working methods, terraforming, geoengineering
- 4. Learning from others…Interdisciplinary Most game-changing developments emerge around the fringes or intersections of disciplines Regular renewal of personnel Regular in-flow of new insights keeps team on the leading edge Encourage taking risks Encourage and reward high risk / high gain activities Scientific rigour and competence Avoid drifting into the realm of science fiction Support from top-management Activities tend to be ridiculed, admired, not taken seriously or seen as threat to core of the establishment. Close ties with academia Most relevant ideas/concepts on a time horizon of 10+ years are generated within academia and research centres
- 5. ACT Research AreasFundamental Physics Impact of new ideas in physics on the space sector Biomimetics & Bioengineering Benefitting from Darwinian evolution to solve engineering problems Mission Analysis Mathematical techniques for future mission analysis Artificial Intelligence Engineering of intelligent computer systems Advanced Energy Systems Innovating energy systems Planetary System Science Options and opportunities from complex climate systems Computer Science & Applied Mathematics Post von neumann architectures Advanced Propulsion Explore and review break- through propulsion concepts Computational Management Science Explore computational aspects of management Advanced Materials Benefitting from the control at micro/nano scale
- 6. We are currently hiring 5 new Research fellows (post-docs)! 1 - Artificial Intelligence 2 - Computer science 3 - Biomimetics 4 - Fundamental Physics 5 - Mission Analysis Deadline 6th July! www.esa.int/act
- 7. 2040 How intelligent will satellites be?
- 8. 80486 Before looking 25 ahead, let us look 25 back at AI and CS
- 10. Today19901965
- 11. 19901970 Today
- 12. And what about algorithms?
- 13. Alpha GO 2016 2000 CNN, BP, LSTM, RNN ImageNet 2006 2012 • 1980-1990: attempts to train DNNs failed • 2006: first worldwide success stories in 2006 Deep Belief Networks and autoencoders: networks trained layer by layer • 2006-2016: great success and explosion of DL, for example: Convolutional Neural Networks (CNNs): ImageNet success Long Short-Term Memory (LSTM): huge success in speech recognition . Just A Hype? No, DL is here to stay. Deep Learning First DL success
- 14. Genetic Programming Symbolic regression (SR): Learn the underlying physics from data Symbolic regression leverages an “evolutionary” approach to model creation, testing billions of potential models per second, and converging on the simplest, most accurate ones that explain your data. S.R. makes no prior assumptions about the data set, instead fitting models to the data dynamically. Schmidt M., Lipson H. (2009) "Distilling Free-Form Natural Laws from Experimental Data," Science, Vol. 324, no. 5923, pp. 81 - 85. Companies using Nutonian SR tool: Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR ImageNet 2006 2012 2009Nutonian First DL success
- 15. Smart “search” (optimization) methods Evolutionary algorithms: exploiting artificial selection to evolve increasingly better solutions to design problems Orders of magnitude better from Genetic Algorithm (80s) to modern techniques: Covariant Adaptation Evolutionary Strategy (CMA-ES), Multi-objective Evolutionary techniques via decomposition (MOEA/D) and Self-adaptive Differential Evolution (jDE) Monte Carlo Tree Search: for sequential decision-making problems One of the most successful techniques in AI for games Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009Nutonian First DL success, jDE 2003CMA-ES 2007MOEA/D
- 16. Perception, understanding and communication Sensors: Dynamic Vision Sensors (DVS) Elementary Motion detectors (EMD) Light Field cameras (LFC) ... Algorithms: SIFT - Scale-invariant feature transform CNN - Convolutional Networks Using TTC, OF - Time to contact, optic flow LSTM - Long Short Term Memory Networks Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009Nutonian A LSTM wins ICDAR handwriting 2003CMA-ES 2007MOEA/D DVS 2013 EMD, SIFT LFC from Stanford 2004 First DL success, jDE
- 17. What did all this progress buy us?
- 18. 1. Text recognition 2. Colorization of black and white images 3. Adding sounds to silent movies 4. Object classification and detection in photographs 5. Generate image from caption 6. Handwriting generation 7. Text Generation (scripts, poetry, etc.) 8. Image Caption Generation 9. Music Composition 10. Software continuous integration 11. Manage currencies 12. Drive cars 13. Navigate 14. Chat 15. Generative Design Some jobs computers can perform (and that could not 25 years ago)
- 19. Application Areas of AI Self-driving vehicles Google, Tesla, Mercedes-Benz, etc. Autonomous flying (drones) Amazon Prime Air delivery Military Drones Robotics Factory automation, Medicine, Scientific exploration ...
- 20. Application Areas of AI Virtual Assistants Cortana, Siri, Viv Language-based services Machine translation Document summarization Emotionally aware interfaces Affective computing
- 21. The next big things in AI/CS (10-20 years ahead)
- 22. 100,000,000,000 1,000,000,000,000 10,000,000,000,000 2040 100,000,000,000,000 Number of neurons in the human brain (~2025) GV100 Volta (NVIDIA GPU)
- 23. In the same place as where ANNs were in the 90s, these technologies hold great potential, and may become the next big things ● Artificial Evolution (Evolutionary Computing) -> Designing the unexpected ● Genetic Programming -> Computers programming themselves ● Artificial Life -> Digital ecologies tHE nEXT bIG tHINGS are today’s “failures” The seeds of these innovations are well planted
- 24. The 2006 NASA ST5 spacecraft antenna (found by Genetic Programming) The ST5 mission successfully launched on March 22, 2006, and so this evolved antenna represents the world's first artificially-evolved object to fly in space
- 25. The ESA (ACT) VLBI GTOC8 trajectory In 2011 the Humies Gold Medal Award was awarded to the ACT work on “Search for a grand tour of the Jupiter Galilean moons” for human-competitive results that were produced by any form of genetic and evolutionary computation.
- 27. R3000: New Horizons RAD6000: Spirit-Opportunity, Messenger, Deep Space 1, Dawn RAD750: Kepler, Juno, Curiosity i386: ISS x86: Falcon 9, Hubble The Excuses: Radiation Tolerance Reliability Satellite build time Launch delays Paperwork Power Consumption NGMP (ESA, LEON4)
- 28. Scenario #1: the gap is not filled. in 2040 the intelligence on board spacecraft will feel as exciting as a videogame from the 90s
- 29. Scenario #2: the gap is filled. in 2040 the intelligence on board spacecraft will compare to today’s situation as modern VR based games compare to Pong
- 30. ACT and AI research
- 31. Explored areas – Neurocontrollers Evolution in robotic islands: ALife in the Galapagos Deep Reinforcement learning for Spacecraft hovering near unkown small bodies Morphological evolution of soft robots at different gravity levels Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009 2003CMA-ES 2007MOEA/D DVS 2013 EMD, SIFT LFC from Stanford 2004 First DL success, jDE 2009Nutonian A LSTM wins ICDAR handwriting
- 32. Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009 2003CMA-ES 2007MOEA/D DVS 2013 EMD, SIFT LFC from Stanford 2004 First DL success, jDE Explored areas – Swarm Intelligence Decentralized Formation Flight with collision avoidance: Equilibrium Shaping Autonomous self-assembly of large space structures Root Swarm: Sensor webs deployment ACT MIT SPHERES experiments: first ANN controlling multiple (homogeneous) agents in space Nutonian A LSTM wins ICDAR handwriting
- 33. Optic flow based lunar landing: from bees to Apollo Scent of science: from a female chasing moth to the chase of methane on Mars Explored areas – Biomimetic Sensing and Actuation Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009 2003CMA-ES 2007MOEA/D DVS 2013 EMD, SIFT LFC from Stanford 2004 First DL success, jDE
- 34. Explored areas – Vision Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009 2003CMA-ES 2007MOEA/D DVS 2013 EMD, SIFT LFC from Stanford 2004 First DL success, jDE Nutonian A LSTM wins ICDAR handwriting Astro Drone - gamification for the acquisition of vision data-sets Learning “to see” in zero gravity - from stereo vision to monocular vision (using the MIT SPHERES platform)
- 35. Explored areas – Evolution and smart search Evolution of Interplanetary Trajectories Parallel evolution in modern CPU architectures, the island model, PyGMO Novel tree search paradigms: Monte Carlo Tree Search, Ant Colony Optimization, Lazy Race Tree Search Humies Gold medal - “for Human-Competitive Results Produced by Genetic and Evolutionary Computation” 1st place in the 8th edition of GTOC - “The America’s cup of rocket science” Alpha GO 2016 2000 CNN, BP, LSTM, RNN, GP, SR, DE, MCTS ImageNet 2006 2012 2009 2003CMA-ES 2007MOEA/D DVS 2013 EMD, SIFT LFC from Stanford 2004 First DL success, jDE Nutonian A LSTM wins ICDAR handwriting
- 36. The scikit-learn of evolutionary algorithms
- 37. pip install pagmo conda config --add channels conda-forge conda install pagmo ● Provides “free” parallelization via the asynchronous island model ● mpi, threads, multiprocess, etc.. all encapsulated in the island ● available for osx, linux and windows ● Fully FLOSS philosophy ● Easily extendible with your own algorithms or problems ● Tutorials and doc constantly up to date ● Community support active via a dedicated gitter channel https://esa.github.io/pagmo2/index.html pagmo/pygmo 2.x
- 38. whats ahead? Differentiable intelligence Use of high order derivative information in ML
- 39. Background: the algebra of floating points >>> def my_function(x): ... return cos(x[0])+(x[0]+3*x[1]+x[2])**7 >>> x = [0.1,0.2,0.3] >>> my_function(x) 1.9896041652780259 Behind this seemingly trivial computation, a number of implicit assumptions we tend to forget. Note: we rarely question that the floating point algebra is “conformal” to the real number algebra.
- 40. Background: the algebra of Truncated Taylor polynomials >>> def my_function(x): ... return cos(x[0])+(x[0]+3*x[1]+x[2])**7 >>> x = [gdual(0.1,"x0",5), gdual(0.2,"x1",5), gdual(0.3,"x2", 5)] >>> my_function(x) 42*dx0*dx2+105*dx0*dx2**2+630*dx0*dx1*dx2+126*dx0*dx1+945*dx0*dx1**2+3 5.0166*dx0**3+6.90017*dx0+20.82*dx2**2+6.946*dx2+125.73*dx1*dx2+314.1* dx1*dx2**2+105*dx0**2*dx2+34.8*dx2**3+945*dx1**2*dx2+945*dx1**3+20.502 5*dx0**2+189*dx1**2+20.973*dx1+ 1.9896+315*dx0**2*dx1 Substitute the algebra of floating point numbers with the algebra of Taylor series expansion, trust it to be conformal to the algebra of continuous functions -> a differential Algebra [4].
- 43. floats floats Hi! I am a computer program
- 44. Any-order Taylor expansion of the program outputs floats floats Hi! I am a computer program
- 45. Application to Machine learning and Evolutionary Computations Gradient: widely used (backprop, SQP, interior point) Hessian: less used, often approximated, anyway researched Traditionally the error (or fitness) is conceived as a real number: instead, consider it as a function (of whatever parameters you choose). Using the new algebra, represent it, in the computer, as a truncated Taylor polynomial just as before you were representing it as a floating point.
- 46. Application to Machine learning and Evolutionary Computations Traditionally the error (or fitness) is conceived as a real number: instead, consider it as a function (of whatever params you choose). Using the new algebra, represent it, in the computer, as a truncated Taylor polynomial just as before you were representing it as a floating point. Similar complexity as Hessians and gradients, but an entirely unresearched field both in machine learning and evolutionary computations.
- 47. Deep learning in deep space Experiments on landing
- 48. Status quo: Precompute a Reference optimal trajectory
- 49. Status quo: Linearize around it to account for disturbancies
- 50. Status quo: Risk to get unstable behaviour if we go off nominal
- 51. Status quo: Use polynomials and perform regression Risk to be suboptimal
- 52. 1. Pre-compute many optimal trajectories 2. Train an artificial neural network to approximate the optimal behaviour 3. Use the network to drive the spacecraft Our approach:
- 53. Quadrotor state: [x, vx , z, zx , θ] control:[u1 , u2 ] Spacecraft state: [m, x, vx , z, zx , θ] control:[u1 , u2 ] state: [m, x, vx , z, vz , θ, vθ ] control: [u1 , u2 , u3 ] Rocket state: [m, x, vx , z, vz , θ, vθ ] control:[u1 , u2 ] Some landing models:
- 54. Goal: Solve the deterministic continuous-time optimal control problem, that is: Current methods (direct or indirect) are not suitable for real-time on-board implementation, an alternative is to correct the deviations from a precomputed profile or use polynomial fits. Hamilton-Jacobi-Bellman equation 1 - precompute many optimal solutions:
- 55. Direct methods Indirect methods - Hermite-Simpson transcription and non-linear programming (NLP) solver - Fast and easy implementation - Chattering effects in the training data have a huge negative impact on the results. - regularization techniques are used to remove them. - Suboptimal results - Solve the Hamilton-Jacobi- Bellman equations with shooting methods - Provides the actual optimal trajectories - But… an initial guess is necessary to solve the problem and it is really difficult to find them - More difficult and awkward implementation 1 - precompute many optimal solutions:
- 56. Two methods to generate the data 1 - precompute many optimal solutions:
- 57. Optimization for different problems: Free Landing - Pinpoint Landing, Time optimal - Power optimal - Mass optimal Resulting in different control profiles: 1 - precompute many optimal solutions:
- 58. - The initial state of each trajectory is randomly selected from a training area - 150,000 trajectories are generated for each one of the problems - Computing the trajectory for a specific starting point is difficult, but we speed up the process of generating random optimal trajectories by: Random walks Homotopy methods 1 - precompute many optimal solutions:
- 59. - The networks are trained on the state-control action paris of the trajectories - Networks with 1 - 5 hidden layers - Supervised Learning - Trained with Stochastic Gradient Descent (and momentum) - Minimize the squared loss error (C) 2 - Approximate state-action with a DNN
- 60. ● Deep networks are always better than shallow networks with the same number of parameters. Challenging landing example 2 - Approximate state-action with a DNN
- 61. DNNs with supervised learning and large datasets successfully approximate the optimal control 2 - Approximate state-action with a DNN
- 62. DNNs with supervised learning and large datasets successfully approximate the optimal control 2 - Approximate state-action with a DNN
- 63. Very accurate results + The DNNs can be used as an on-board reactive system (32.000x faster than optimization methods used to generate the data) 3 - How good is it?
- 64. ● Successful landings from states outside of the training initial conditions ● This suggest that an approximation to the solution of the HJB equations is learned Multicopter (power) Multicopter (power) Spacecraft I generalization
- 66. After reaching the target point the spacecraft hovers until it runs of fuel Is it learning the dynamics of the model? No training data below this line generalization
- 67. The system was evaluated with the Parrot Bepop Drone in the TU Delft University Optimal trajectory and trajectory followed by the drone (after some scaling to adjust them, don’t trust this image) Optimal trajectories generated for the Bepop drone 4 - thE REAL WORLD
- 68. 4 - thE REAL WORLD
- 69. Adding a cnn for the perception
- 70. 1 - Train a neural network to guess the state from an on-board camera 2 - Use it together with the previous DNNs to get fully automated visual landing CNN for state estimation from camera
- 71. A simple setup is used: a 3D model (Blender) of a rocket landing on a sea platform (Falcon 9 inspired) CNN for state estimation from camera
- 72. CNN for state estimation from camera
- 73. CNN for state estimation from camera
- 74. CNN for state estimation from camera
- 76. References: [1] Carlos Sánchez-Sánchez, Dario Izzo and Daniel Hennes. "Optimal real-time landing using deep networks." Proceedings of the Sixth International Conference on Astrodynamics Tools and Techniques, ICATT. Vol. 12. 2016. [2] Carlos Sánchez-Sánchez, Dario Izzo and Daniel Hennes. "Learning the optimal state-feedback using deep networks." Computational Intelligence (SSCI), 2016 IEEE Symposium Series on. IEEE, 2016. [3] Carlos Sánchez-Sánchez and Dario Izzo. "Real-time optimal control via Deep Neural Networks: study on landing problems." arXiv preprint arXiv:1610.08668(2016).
- 78. Encoding a computer program weighted Cartesian Genetic Program Miller, Julian F., and Peter Thomson. "Cartesian genetic programming." European Conference on Genetic Programming. Springer Berlin Heidelberg, 2000.
- 79. weighted dCGP Case B: all floats except w3,1 and w10,1 -> gduals operating in P3 : Case A: all floats The output is a float The output is a Taylor polynomial in w3,1 and w10,1 truncated at the third order
- 80. Weighted CGP >>> ks = kernel_set(["sum","diff","mul","div"])() >>> CGP = expression(1,1,1,10,10,2,ks, seed = 21312312) >>> print(CGP(["x"])) [x**2 - x] >>> print(CGP([0.1])) [-0.09] >>> ks = kernel_set(["sum","diff","mul","div"])() >>> CGPw = expression(1,1,1,10,10,2,ks, seed = 21312312) >>> print(CGPw(["x"])) [w1_0*w1_1*w2_0*x**2 - w2_1*x] Traditional CGP: all floats, no weights Case A: all floats, weighted expression
- 81. Weighted dCGP Case B: define w1_0, w2_0 and w2_1 as gduals, add also the input >>> ks = kernel_set(["sum","diff","mul","div"])() >>> dCGPw = expression(1,1,1,10,10,2,ks, seed = 21312312) >>> print(dCGPw(["x"])) [w1_0*w2_0*x**2 - w2_1*x] >>> ex.set_weight(1, 0, gdual(1., "w1_0", 2)) >>> ex.set_weight(2, 0, gdual(1., "w2_0", 2)) >>> ex.set_weight(2, 1, gdual(1., "w2_1", 2)) >>> print(dCGPw([gdual(0.1)])) [-0.1*dw2_1+0.01*dw1_0*dw2_0+0.01*dw2_0-0.09+0.01*dw1_0] >>> print(dCGPw([gdual(0.1, “x”, 2)])) [0.01*dw1_0*dw2_0+dw1_0*dx0**2+0.2*dw1_0*dw2_0*dx0+0.2*dw1_0*dx0+0.01*d w1_0+dx0**2-0.1*dw2_1+0.2*dw2_0*dx0+0.01*dw2_0-0.09-0.8*dx0-dw2_1*dx0+d w2_0*dx0**2]
- 84. Example 1: Learning ephemeral constants Taking the skeleton out of the closet …. ‘‘...the finding of numerical constants is a skeleton in the GP closet...[and an] area of research that requires more investigation...’’ - John Koza
- 85. 1 - The error of a CGP expression is computed in the new algebra (so it’s a function, remember….). 2 - We thus get the order n Taylor expansion of the error: example with one only constant and order three 3 - We use the differential expression obtained for the error to update the ephemeral constant values so that the error is minimized 4 - At order 1 and with some type of gradient descent, you may think of this as a backpropagation to learn ephemeral constant values. Example 1: Learning ephemeral constants
- 86. What order? Using: Expressions such as: Result in the error: A parabolid! -> A second order Taylor polynomial represents this error exactly.
- 87. The final algorithm (1-4)-ES evolves the chromosome and thus the symbolic expression y(x,c). We assign to each fitness the minimum of the MSE across all possible values of the constants: The solution is approximated by one step of the Newton method. Hessian and gradient are extracted from a second order Taylor approximation
- 88. Learning ephemeral constants: results ● Success: MSE < 1e-14 (i.e. we learn the exact value of the constants) ● We sample ~50 points in a uniform grid within bounds. ● We perform 100 runs. ● We compute the Expected Run Time (ERT): the expected value of the number of d-CGP expressions that have to be evaluated before meeting the success criteria set. ● Closest work is Topchy and Punch [1]: not comparable to these results.
- 89. Example 2: Learning constants with weighted dCGP Destroying the closet
- 90. Weight batch learning (1-4)-ES evolves the chromosome and thus the symbolic expression y(x,c). We assign to each fitness the minimum of the MSE across all possible values of the weights: Solution by the Newton method is now troublesome. A new learning method: weight batch learning We basically perform one Newton step to learn a batch of two weights at a time. No Lamarckian learning: at each generation weights are sampled from a normal distribution.
- 91. Learning constants with weighted dCGP: results ● Same setup as in the previous experiments. ● All constants are learned within the set precision. ● Generally requires less generations of the ES ● ERT is higher because of the Newton iterations.
- 92. Example 3: Solving differential equations Cauchy problems, Neumann and Dirichlet problems, TPBV problems
- 93. Solving Differential Equations: simple example with y(0.1)=20.1 and x in [0.1,1] We assume is represented by our d-CGP expression (1 in 1 out). We construct a grid of 10 values for x between 0.1 and 1. Computing x as a gdual with truncation order 1, we get from the Taylor expansion of the program output. Following Tsoulos and Lagaris [2], at each generation we use as error the sum of two terms: - The violation of the differential equation: - The violation of the boundary condition:
- 94. Solving Differential Equations: results Problem ERT dCGP ERT Tsoulos 8123 130600 35482 148400 22600 88200 896 38200 24192 40600 327020 797000 Comparison to Tsoulos and Lagaris [2] work possible. CGP outperforming grammatical evolution in these tasks d-CGP generalizing Tsoulos and Lagaris [2] method allowing high order and mixed derivatives
- 95. Example 4: Finding prime integrals From differential equations to the fundamental conservation laws
- 96. Finding prime integrals Prime integrals are typically found by great mathematicians and their intuition dynamical system in normal form a prime integral
- 97. Swinging Atwood Machine Not only energy … “whatever this is” … is conserved
- 98. Found in 1888 as a third example of integrable top Kovalevskaya Top Not only angular momentum … “whatever this is” … is conserved
- 99. Finding prime integrals No need to solve this to create training data (i.e. no need to observe the system as in Schmidt and Lipson [3]) We get the derivatives from the 1st order Taylor expansion of the program output!
- 100. Finding prime integrals: results
- 101. References: [1] Topchy, Alexander, and William F. Punch. "Faster genetic programming based on local gradient search of numeric leaf values." Proceedings of the 3rd Annual Conference on Genetic and Evolutionary Computation. Morgan Kaufmann Publishers Inc., 2001. [2] Tsoulos, Ioannis G., and Isaac E. Lagaris. "Solving differential equations with genetic programming." Genetic Programming and Evolvable Machines 7.1 (2006): 33-54. [3] Schmidt, Michael, and Hod Lipson. "Distilling free-form natural laws from experimental data." Science 324.5923 (2009): 81-85. [4] Ritt, Joseph Fels. Differential algebra. Vol. 33. American Mathematical Soc., 1950. [5] Izzo, Dario, Francesco Biscani, and Alessio Mereta. "Differentiable Genetic Programming." European Conference on Genetic Programming. Springer, Cham, 2017.
- 102. The kelvins platform Compete to learn and train
- 103. Kelvins Portal: compete to excel • Asking the correct questions is essential to be successful in science. • A dedicated competition portal: Kelvins, reach the absolute zero (error). • AlgoritHmic and data mining competitions co-exist • Targeting machine learning, data mining communities but also space engineers. • Portal: https://kelvins.esa.int/ • Competitions in the pipeline: Asteroid belt / debris surrogate models, orbital propagation error prediction.
- 104. Mars Express Power Challenge • predict the power consumption of the spacecraft thermal subsystem. • Three years of spacecraft telemetry are released … can you predict the fourth year? • The ultimate goal is to automate operations and extend satellite life time, which in turn increases the scientific return.
- 105. Fact sheet ● Downloads: 650 ● Number of different countries > 20 ● Teams in the final leaderboard: 40 ● Registered teams: 133 ● Submitted solutions: ~200
- 106. Winners Jozef Stefan Institute, Lujbjana, Slovenia Codename: MMMe8 Department Of Knowledge Technologies: Prof. Saso Dzeroski SCORE (RMSE lower is better) 0.07916
- 107. Runner-up Stephanos Stephani Codename: redrock github.com/stephanos-stephani/MarsExpressChallenge SCORE (RMSE lower is better) 0.08030
- 108. Final results and public leader-board
- 109. Evolution of the scores and events Competition time Starter Kit RMSE Score Fair Perfect
- 116. Star Trackers: First Contact (ongoing) • A spacecraft is lost in space and needs to autonomously determine its attitude based on the camera image of a star tracker. • Given 10 000 images of such a scenario, participants of the competition have to identify stars visible in the images. • The goal is to improve state-of-the-art algorithms in terms of speed, accuracy and robustness
- 117. The Kessler Run: GTOC9 It is the year 2060 and the commercial exploitation of Low Earth Orbits (LEOs) went well beyond the trillion of Euros market size. Following the unprecedented explosion of a Sun-synchronous satellite, the Kessler effect triggered further impacts and the Sun-synchronous LEO environment was severely compromised. Scientists from all main space agencies and private space companies isolated a set of 123 orbiting debris pieces that, if removed, would restore the possibility to operate in that precious orbital environment and prevent the Kessler effect to permanently compromise it. You are thus called to design a series of missions able to remove all critical debris pieces while minimizing the overall cumulative cost of such an endeavour. Each single mission cost (in EUR) will depend on how early the mission is submitted via this web-site (regardless of their actual launch epoch) and on the spacecraft initial mass.
- 118. The Kessler Run: GTOC9 ● Number of different countries: 19 ● Teams in the final leaderboard: 36 ● Registered teams: 69 ● Registered institutions: 125 ● Scientists registered: ~320 ● Missions submitted:~1200 ● A difficult combinatorial problem with complex optimization procedures to evaluate the various heuristics / costs involved in transfers. ● Some Links to tsp variants or set cover. ● Won by an approach based on genetic algorithms and ant colony optimization by Jet propulsion Laboratory. ● Surrogate models suggested and used successfully.
- 119. Interplanetary Trajectory Planning with Monte Carlo Tree Search Hennes, Daniel, and Dario Izzo. "Interplanetary trajectory planning with Monte Carlo tree search." Proceedings of the 24th International Conference on Artificial Intelligence,” AAAI Press. 2015. http://ijcai.org/Proceedings/15/Papers/114.pdf Optimal real-time landing using deep networks Sánchez-Sánchez, Carlos, Dario Izzo, and Daniel Hennes. "Optimal Real-Time Landing Using Deep Networks." http://www.esa.int/gsp/ACT/doc/AI/pub/ACT-RPR-AI-2016-ICATT-optimal_landing_deep_networks.pdf Evolving solutions to TSP variants for active space debris removal. Izzo, Dario, et al. "Evolving solutions to TSP variants for active space debris removal." Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation. ACM, 2015. An evolutionary robotics approach for the distributed control of satellite formations Izzo, Dario, Luís F. Simões, and Guido CHE de Croon. "An evolutionary robotics approach for the distributed control of satellite formations."Evolutionary Intelligence 7.2 (2014): 107-118. Search for a grand tour of the jupiter galilean moons Izzo, Dario, et al. "Search for a grand tour of the jupiter galilean moons." Proceedings of the 15th annual conference on Genetic and evolutionary computation. ACM, 2013. Evolutionary robotics approach to odor source localization De Croon, G. C. H. E., et al. "Evolutionary robotics approach to odor source localization." Neurocomputing 121 (2013): 481-497. Novelty search for soft robotic space exploration Methenitis, Georgios, et al. "Novelty search for soft robotic space exploration." Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation. ACM, 2015. Lattice formation in space for a swarm of pico satellites Pinciroli, Carlo, et al. "Lattice formation in space for a swarm of pico satellites." International Conference on Ant Colony Optimization and Swarm Intelligence. Springer Berlin Heidelberg, 2008. Selected ACT bibliography (more here)