Asteroseismologie spatiala (participarea la misiunile spatiale KEPLER, CoRoT, MOST)

si tehnici de supercalcul in astronomia

romaneasca

Marian Doru Suran, Dumitru Pricopi

Astronomical Institute of the Romanian Academy

E-mail: suran@aira.astro.ro , dpricopi@aira.astro.ro

Supercomputers/supercomputingSupercomputer• A supercomputer is a computer that is at the frontline of current processing capacity,

particularly speed of calculation. (en.wikipedia.org/wiki/Supercomputing)

• A supercomputer is a computer that is especially designed to receive, process and present very large amounts of data very quickly. (www.spaceday.org/index.php/Glossary-of-Aeronautics-Terms.html)

• Supercomputers were introduced in the 1960s and were designed primarily by Seymour Cray at Control Data Corporation (CDC), which led the market into the

1970s until Cray left to form his own company, Cray Research.

Supercomputing• Supercomputers are used for highly calculation-intensive tasks such as problems

involving quantum physics, weather forecasting, climate research, molecular modeling (computing the structures and properties of chemical compounds, biological macromolecules, polymers, and crystals), physical simulations (such as simulation of airplanes in wind tunnels, simulation of the detonation of nuclear weapons, and research into nuclear fusion).

• A particular class of problems, known as Grand Challenge problems, are problems whose full solution requires semi-infinite computing resources.

SupercomputingGrand Challenge Problems• Grand Challenges were USA policy terms set as goals in the late 1980s for funding high-

performance computing and communications research in part in response to the Japanese 5th Generation (or Next Generation) 10-year project.

• "A grand challenge is a fundamental problem in science or engineering, with broad applications, whose solution would be enabled by the application of high performance computing resources that could become available in the near future. Examples of grand challenges are: Computational fluid dynamics for

the design of hypersonic aircraft, efficient automobile bodies, and extremely quiet submarines, weather forecasting for short and long term effects, efficient recovery of oil, and for many other applications;

Electronic structure calculations for the design of new materials such as chemical catalysts, immunological agents, and superconductors;

Plasma dynamics for fusion energy technology and for safe and efficient military technology; Calculations to understand the fundamental nature of matter, including quantum chromodynamics and

condensed matter theory; Symbolic computations including

Speech/pattern recognition, global search heuristics computer vision, natural language understanding, automated reasoning, global search heuristics, tools for design, manufacturing, and simulation of complex systems."

["A Research and Development Strategy for High Performance Computing", Executive Office of the President, Office of Science and Technology Policy, November 20, 1987]

Supercomputing• Executive Office of the President, Office of Science and Technology Policy, "The Federal

High Performance Computing Program," Sept. 1989, pp. 49–50: Appendix A Summary:

• Prediction of weather, climate, and global change • Challenges in materials sciences • Semiconductor design • Structural biology • Design of pharmaceutical drugs • Human genome • Quantum Chromodynamics • Astronomy• Challenges in Transportation • Vehicle Signature • Turbulence • Vehicle dynamics • Nuclear fusion • Efficiency of combustion systems • Enhanced oil and gas recovery • Computational ocean sciences • Speech • Vision • Undersea surveillance for ASW

SupercomputingAstronomical Institute Rom. Acad.

(AIRA)

• Computational fluid dynamics: MHD HD Nbody+SPH.

• Calculations to understand the fundamental nature of matter Universe formation and evolution/from the inflation period to now.

• Symbolic computations including: Recognition processes, Automated reasoning, Global search heuristics

Artificial intelligence

SupercomputersModern supercomputer architecture

Within this hierarchy we have:• A multiprocessing computer is a computer, operating under a single OS and using more than one CPU, wherein the application-

level software is indifferent to the number of processors. The processors share tasks using Symmetric multiprocessing (SMP) and Non-Uniform Memory Access (NUMA).

• A computer cluster is a collection of computers that are highly interconnected via a high-speed network or switching fabric. Each computer runs under a separate instance of an Operating System (OS).

• A SIMD processor executes the same instruction on more than one set of data at the same time. The processor could be a general purpose commodity processor or special-purpose vector processor. It could also be high-performance processor or a low power processor. As of 2007, the processor executes several SIMD instructions per nanosecond.

Supercomputer challenges, technologies, problems

• A supercomputer generates large amounts of heat and must be cooled. Cooling most supercomputers is a major HVAC problem. • Information cannot move faster than the speed of light between two parts of a supercomputer. For this reason, a supercomputer that is

many metres across must have latencies between its components measured at least in the tens of nanoseconds. Seymour Cray's supercomputer designs attempted to keep cable runs as short as possible for this reason, hence the cylindrical shape of his Cray range of computers. In modern supercomputers built of many conventional CPUs running in parallel, latencies of 1–5 microseconds to send a message between CPUs are typical.

• Supercomputers consume and produce massive amounts of data in a very short period of time. According to Ken Batcher, "A supercomputer is a device for turning compute-bound problems into I/O-bound problems." Much work on external storage bandwidth is needed to ensure that this information can be transferred quickly and stored/retrieved correctly.

Technologies developed for supercomputers include:

• Vector processing • Liquid cooling • Non-Uniform Memory Access (NUMA)• Striped disks (the first instance of what was later called RAID) • Parallel filesystems s• video game consoles (graphics processing units (GPUs))

NUMA SupercomputersNon-Uniform Memory Access or Non-UniformMemory Architecture (NUMA):

is a computer memory design used in multiprocessors, where the memory access time depends on the memory location relative to a processor. Under NUMA, a processor can access its own local memory faster than non-local memory, that is, memory local to another processor or memory shared between processors.

• NUMA architectures logically follow in scaling from symmetric multiprocessing (SMP) architectures. Their commercial development came in work by Burroughs (later Unisys), Convex Computer (later Hewlett-Packard), Silicon Graphics, Sequent Computer Systems, Data General and Digital during the 1990s. Techniques developed by these companies later featured in a variety of Unix-like operating systems, and somewhat in Windows NT.

One possible architecture of a NUMA system. Notice that the processors are connected to the bus or crossbar by connections of varying thickness/number. This shows that different CPUs have different priorities to memory access based on their location.

SGI Supercomputers

Silicon Graphics (SGI) supercomputers:

• Silicon Graphics, Inc. (commonly initialised to SGI, historically sometimes referred to as Silicon Graphics Computer Systems or SGCS) was a manufacturer of high-performance computing solutions, including computer hardware and software, founded in 1981 by Jim Clark and Abbey Silverstone. Its initial market was 3D graphics display terminals, but its products, strategies and market positions evolved significantly over time.

Switch to Itanium• In 1998, SGI announced that future generations of its machines would be based not on their own MIPS processors, but the new “super-

chip” from Intel, Itanium. Funding for its own high-end processors was reduced, and it was planned that the R10000 would be the last MIPS mainstream processor. MIPS would focus entirely on the embedded market, where it was having some success, and SGI would no longer have to fund development of a CPU that, since the failure of ARC, found use only in their own machines.

• But this plan quickly went awry. As early as 1999 it was clear the Itanium was going to be delivered very late, and then that it would have nowhere near the performance originally expected. As the production delays increased, MIPS's existing R10000-based machines grew increasingly uncompetitive. Eventually it was forced to introduce faster MIPS processors, the R12000, R14000 and R16000, which were used in a series of models from 2002 until 2006.

• SGI's first Itanium-based system was the short-lived SGI 750 workstation, launched in 2001. SGI's MIPS-based systems were not to be superseded until the launch of the Itanium 2-based Altix servers and Prism workstations some time later. Unlike the MIPS systems, these models used SuSE Linux Enterprise Server with SGI enhancements as their operating system instead of IRIX. SGI uses Transitive Corporation's QuickTransit software to allow their old MIPS/IRIX applications to run (in emulation) on the new Itanium/Linux platform.

• In the server space the Itanium 2-based Altix eventually replaced the MIPS-based Origin product line. In the workstation space, the switch to Itanium was not completed before SGI exited this market.

• The Altix was the most powerful computer in the world in 2006, if a "computer" is defined as a collection of hardware running under a single instance of an operating system. The Altix had 512 Itanium processors running under a single instance of Linux OS. A cluster of 20 machines was then the eighth fastest supercomputer. All faster supercomputers were clusters, but none have as many FLOPS per machine. However, more recent supercomputers are massive clusters of machines that are individually less capable. SGI acknowledged this and in 2007 moved away from the "massive NUMA" model to efficient clusters.

SGI ALTIX SupercomputersAltix supercomputers (From Wikipedia, the free encyclopedia)

• Altix is a line of servers and supercomputers produced by Silicon Graphics, based on Intel processors. It succeeded the MIPS/IRIX-based Origin 3000 servers.

• The line was first announced[1] on January 7, 2003, with the Altix 3000 series, based on Intel Itanium 2 processors and SGI's NUMAlink processor interconnect. At product introduction, the system supported up to 64 processors running Linux as a single system image and shipped with a Linux distribution called SGI Advanced Linux Environment, which was compatible with Red Hat Advanced Server. By August 2003, many SGI Altix customers[2] were running Linux on 128p and even 256p SGI Altix systems, but SGI officially announced 256-processor support[3] within a single system image of Linux on March 10, 2004 using an 2.4-based kernel. The SGI Advanced Linux Environment was eventually dropped after support using a standard, unmodified SUSE Linux Enterprise Server (SLES) distribution for SGI Altix was provided with SLES 8[4] and SLES 9[5]. Later, SGI Altix 512-processor systems were officially supported[6] using unmodified, standard Linux distribution with the launch of SLES 9 SP1. Besides full support of SGI Altix on SUSE Linux Enterprise Server, a standard and unmodified Red Hat Enterprise Linux was also fully supported starting with SGI Altix 3700 Bx2 with RHEL 4[7] and RHEL 5[8] with system processor limits defined by Red Hat for those releases.

• Altix 3700• The Altix 3700 is a high-end model supporting 16 to 512 processors and 8 GB to 2 TB of memory. It requires one or multiple tall (39U)

rack(s). A variant of the Altix 3000 with graphics capability is known as the Prism.• It is based on the third generation NUMAflex distributed shared memory architecture and it uses the NUMAlink 4 interconnection fabric.

The Altix 3000 supports a single system image of 64 processors. If there are more than 64 processors in a system, then the system must be partitioned.

• The basic building block is the C-brick. Each C-brick contains two nodes. A C-brick is a 4U high rackmount unit. Each node contains two Intel Itanium 2 processors which connects to the Super-Bedrock ASIC through a single front side bus. The Super-Bedrock ASIC is a crossbar for the processors, the local RAM, the network interface and the I/O interface. The two Super-Bedrock ASICs in each brick are connected internally by a single 6.4 GB/s NUMAlink 4 channel. A processor node also contains 16 DIMM slots that accept standard DDR1 DIMMs with capacities of 4 to 16 GB.

• Altix 3700 Bx2• The Altix 3700 Bx2 is a high-end model supporting 16 to 2,048 Itanium 2 processors and 12 GB to 24 TB of memory. It requires

one or multiple tall (40U) rack(s).

SGI ALTIX 3700 Supercomputers

SGI® Altix® 3700 (www.sgi.com)

• A powerful and versatile HPC platform, the Altix 3700 runs standard Linux and delivers breakthrough performance. The newest generation of SGI's high-end servers, Altix 3700 Bx2 continues to win awards for its performance, ease of use and breakthrough capabilities. Its growth is fast outpacing the industry. Altix 3700 also leverages industry standard components to offer the most value in high end computing, starting at around $100,000.

• Altix 3700 is: • The most versatile platform to handle large HPC workloads • Scale up or scale out on one platform—up to 512P in one system • Run small to large jobs in any programming model • Highly efficient development and administration environment • The power behind breakthroughs in science and engineering • Break new barriers with up to 24 TB of addressable memory • Scale even stubborn applications for real sustained performance • Extraordinary "Peer I/O" capability: direct high speed link to memory • The only supercomputing solution fully devoted to Industry standards, including Linux® and

Linux only • Access hundreds of 64-bit technical applications • No-compromise Linux performance with a complete compute and data management solution • Benefit from SGI's leadership in support of HPC in the Linux community • Choice of Novell® SUSE® Linux Enterprise Server or Red Hat® Enterprise Linux® Advanced Server (up

to 64 processors and 64GB memory per system) • Scaling to 512 Itanium 2 processors in a single node, Altix 3700 Bx2 is SGI's newest and most powerful

Linux solution for high performance computing. Altix 3700 Bx2 leverages the powerful SGI® NUMAflex® global shared-memory architecture to derive maximum application performance from new high-density CPU bricks. The latest configuration also doubles available bandwidth between Altix bricks with SGI's NUMAlink® 4 interconnect technology - the industry's fastest at 6.4GB/sec and less than 1 microsecond MPI latency. Each node in an Altix 3700 system can contain 16 to 512 processors, up to 6 terabytes of global shared memory, and 48 XIO™ buses; and delivers over 3 gigabytes per second of sustained I/O bandwidth.

SGI ALTIX 3700 vs.BlueGene Supercomputers

Topology:• The Altix 3000 and 4000 are distributed shared

memory multiprocessors. An Altix 3700 system can contain up to 2048 dual-core Itanium 2 and Itanium ("Montvale" revision) microprocessor sockets, which are connected by the NUMAlink 4 interconnect in a hypercube network topology. The microprocessors are accompanied by up to 128 TB of memory (192TB with single microprocessor socket blades and 16GB DIMMs).

• Each Blue Gene/L node is attached to three parallel communications networks: a 3D network topology for peer-to-peer communication between compute nodes, a for collective communication, and a global interrupt network for fast barriers.

SGI ALTIX 3700 Supercomputers, AIRA

SGI platform in the Astronomical Institute of the Romanian AcademyComputational resources :• 1995 - a supercomputing platform SGI at AIRA (workstations +visualization);• 2007 – June 2008: 4 processors (ALTIX/SGI, Astronomical Institute)• + 8 processors (ALTIX/SGI, SGI test center, US):• June 2008 - : 40 processors (ALTIX/SGI, Astronomical Institute)

The superscalar SGI/ALTIX3700/ITANIUM2 supercomputer:– 44 processors (1.3GHz/3MB Cache2)/shared system;– 250 Gflops computational power; – 2.5 Tb storage;

Environmental specifications :– 60 kBTU cooling system;– 14 kW electrical power (7 kW/ALTIX, 7 kW/cooling sys.)

Supercomputing, AIRA1. High Performance & Scientific Computing:MHD/HD/N-body+SPH• First objects: Chemistry, masses, formation redshifts, Large Scale Structure of the Universe,

Weak lensing effects, Baryon oscillations.• Supernovae: SNe in small halos; OB associations • Star formation: Looking at methods to both infer the number of stars being formed and feed the

energy of star formation back into the simulation. • Galaxy Formation: In both large-scale cosmological simulations and in models of clusters of

galaxies. • X-ray emission Large adiabatic simulations of cluster formation and evolution• Interstellar Medium: S tudies of the formation and evolution of the ISM • Stellar structure: Magnetoconvection+differential rotation. (Super)granulation in stellar atnospheres

2. Global Search Heuristics (Genetic Algorithms)• Stellar Astrophysics Stellar structure modeling using a parallel genetic algorithm

for objective global optimization• Asteroseismology Seismic Inference using Genetic Algorithms

3. Robotic/Thinking Telescope• Robotic and remote control/Thinking 1.3m telescope, AIRA.

MHD/HD/N-body+SPH Supercomputing, AIRA

MHD/HD software

1. High Performance & Scientific Computing:MHD/HD/N-body+SPH. Science Programme

MHD/HD/N-body+SPH Supercomputing, AIRA

N-body software

AI Supercomputing

Artificial inteligence:Artificial intelligence (AI) is the intelligence of machines

and the branch of computer science that aims to create it. Textbooks define the field as "the study and design of intelligent agents,"[1] where an intelligent agent is a system that perceives its environment and takes actions that maximize its chances of success.[2] John McCarthy, who coined the term in 1956,[3] defines it as "the science and engineering of making intelligent machines."[4]

• Machine Learning • Automated Planning

AI supercomputing

2.Machine Learning. Pattern recognition• Pattern recognition is "the act of taking in raw data and

taking an action based on the category of the pattern"[1]. Most research in pattern recognition is about methods for supervised learning and unsupervised learning.

3. Automate reasoning• Automated reasoning is an area of computer science

dedicated to understanding different aspects of reasoning in a way that allows the creation of software which allows computers to reason completely or nearly completely, automatically. As such, it is usually considered a subfield of artificial intelligence, but it also has strong connections to theoretical computer science

AI supercomputingArtificial inteligence:Tools• In the course of 50 years of research, AI has developed a large number of tools to solve the most difficult

problems in computer science. A few of the most general of these methods are discussed below.

2. Classifiers and statistical learning methods• Main articles: Classifier (mathematics), Statistical classification, and Machine learning• The simplest AI applications can be divided into two types: classifiers ("if shiny then diamond") and controllers ("if

shiny then pick up"). Controllers do however also classify conditions before inferring actions, and therefore classification forms a central part of many AI systems. Classifiers are functions that use pattern matching to determine a closest match. They can be tuned according to examples, making them very attractive for use in AI. These examples are known as observations or patterns. In supervised learning, each pattern belongs to a certain predefined class. A class can be seen as a decision that has to be made. All the observations combined with their class labels are known as a data set. When a new observation is received, that observation is classified based on previous experience.[124]

• A classifier can be trained in various ways; there are many statistical and machine learning approaches. The most widely used classifiers are the neural network,[125] kernel methods such as the support vector machine,[126] k-nearest neighbor algorithm,[127] Gaussian mixture model,[128] naive Bayes classifier,[129] and decision tree.[130] The performance of these classifiers have been compared over a wide range of tasks. Classifier performance depends greatly on the characteristics of the data to be classified. There is no single classifier that works best on all given problems; this is also referred to as the "no free lunch" theorem. Determining a suitable classifier for a given problem is still more an art than science.[131]

AI supercomputingArtificial inteligence:Tools• In the course of 50 years of research, AI has developed a large number of tools to solve the most difficult problems in computer science.

A few of the most general of these methods are discussed below.

1. Search and optimization• Main articles: Search algorithm, Optimization (mathematics), and Evolutionary computation• Many problems in AI can be solved in theory by intelligently searching through many possible solutions:[96] Reasoning can be reduced to

performing a search. For example, logical proof can be viewed as searching for a path that leads from premises to conclusions, where each step is the application of an inference rule.[97] Planning algorithms search through trees of goals and subgoals, attempting to find a path to a target goal, a process called means-ends analysis.[98] Robotics algorithms for moving limbs and grasping objects use local searches in configuration space.[67] Many learning algorithms use search algorithms based on optimization.

• Simple exhaustive searches[99] are rarely sufficient for most real world problems: the search space (the number of places to search) quickly grows to astronomical numbers. The result is a search that is too slow or never completes. The solution, for many problems, is to use "heuristics" or "rules of thumb" that eliminate choices that are unlikely to lead to the goal (called "pruning the search tree"). Heuristics supply the program with a "best guess" for what path the solution lies on.[100]

• A very different kind of search came to prominence in the 1990s, based on the mathematical theory of optimization. For many problems, it is possible to begin the search with some form of a guess and then refine the guess incrementally until no more refinements can be made. These algorithms can be visualized as blind hill climbing: we begin the search at a random point on the landscape, and then, by jumps or steps, we keep moving our guess uphill, until we reach the top. Other optimization algorithms are simulated annealing, beam search and random optimization.[101]

• Evolutionary computation uses a form of optimization search. For example, they may begin with a population of organisms (the guesses) and then allow them to mutate and recombine, selecting only the fittest to survive each generation (refining the guesses). Forms of evolutionary computation include swarm intelligence algorithms (such as ant colony or particle swarm optimization)[102] and evolutionary algorithms (such as genetic algorithms[103] and genetic programming[104][105]).

AI SupercomputingPattern recognition• Pattern recognition is "the act of taking in raw data and taking an action based on the category of the pattern"[1]. Most research in

pattern recognition is about methods for supervised learning and unsupervised learning.• Pattern recognition aims to classify data (patterns) based either on a priori knowledge or on statistical information extracted from the

patterns. The patterns to be classified are usually groups of measurements or observations, defining points in an appropriate multidimensional space. This is in contrast to pattern matching, where the pattern is rigidly specified.

Computer data processing• Computer data processing is any process that uses a computer program to enter data and summarise, analyse or otherwise

convert data into usable information. The process may be automated and run on a computer. It involves recording, analysing, sorting, summarising, calculating, disseminating and storing data. Because data is most useful when well-presented and actually informative, data-processing systems are often referred to as information systems. Nevertheless, the terms are roughly synonymous, performing similar conversions; data-processing systems typically manipulate raw data into information, and likewise information systems typically take raw data as input to produce information as output.

• Data processing may or may not be distinguished from data conversion, when the process is merely to convert data to another format, and does not involve any data manipulation.

Statistical classification• Statistical classification is a supervised machine learning procedure in which individual items are placed into groups based on

quantitative information on one or more characteristics inherent in the items (referred to as traits, variables, characters, etc) and based on a training set of previously labeled items.

Evolutionary algorithm• Evolutionary algorithm (EA) is a subset of evolutionary computation, a generic population

Robotics• Robotics is the engineering science and technology of robots, and their design, manufacture, application, and structural disposition.

Numerical control• Numerical control (NC) refers to the automation of machine tools that are operated by abstractly programmed commands encoded on a

storage medium, as opposed to manually controlled via handwheels or levers, or mechanically automated via cams alone.

Supercomputers - my life experiencesYear Type Software, Astronomy Speed Memory

1973-1975

Math. Fac, Bucharest

IBM 360

Scalar, mainframe

MHD 3D, Solar Wind ~250kFLOPS ~8 Mb/8b

DOS/F64

1975-1977

ASE, Bucharest

IBM 370

Scalar, mainframe

MHD 3D, Solar Wind ~1 MFLOPS ~100 Mb/32b

DOS/F64

1977-1990

IFA, Bucharest

IBM 370

Scalar, mainframe

Stellar structure,

Stellar evolution

~1 MFLOPS ~100 Mb/32b

DOS/F64

1987-1990

IFA, Pitesti

CDC Cyber

Superscalar, supercomputer,

Stellar structure,

Stellar evolution

~400 MFLOPS ~100?/16b

DOS/F64

1991-1993

AIRA, Bucharest

IRIS 2000

Scalar, mini

Stellar structure,

Stellar evolution

Cosmology, N-body

~50 Mflops ~10 MB/32b

DOS/F64

1994-1998,

Paris-Meudon Obs.

CONVEX

Vector, supercomputer

Stellar evolution, stellar pulsation,

asteroseismology

~50Mflops/proc. UX/64b

/F77

1995-

AIRA, Bucharest

SGI Power Challenge

Superscalar, workstation

1 proc.

Stellar structure,

Stellar evolution

Cosmology, N-body+SPH

~360 MFLOPS 768 Mb/64b

IRIX/F77, F90

2003 –

AIRA, Bucharest

SGI ALTIX 330/

4 proc

Superscalar

supercomputer

Stellar structure,

Stellar evolution

Cosmology, N-body+SPH,

Neural Net, Genetic alg.

~24 GFLOPS 12 Gb/64b

REDHAT INTER.

Intel Fortran

2008 –

AIRA, Bucharest

SGI ALTIX 3700

40 proc

Superscalar

supercomputer

Stellar evolution

Cosmology, N-body+SPH,

Neural Net, Genetic alg.

Thinking telescope?

~240 GFLOPS 80 Gb/64b

SUSE Interp./

Intel Fortran

Supercomputing, AIRA

Results

1.a. 1.a. ΛΛCDM Hydrodynamical CDM Hydrodynamical Cosmological Simulations, AIRACosmological Simulations, AIRA

HD Cosmological simulations, AIRA

High z cosmological simulations Cosmological scenario: ΛCDM (b [SPH], dm,[N-body] , Λ, h) Cosmological model: WMAP5+BAO+SN; = 0.726 , DM =0.208 , b =0.045 , h2 = 0.0071, = 1.00, h = 70.5 ;• numerical simulation N-body type Art-PM; P3M+SPH;• high z: 12 z 5, avoiding reionization (stellar feedback);Computing:• L=100 h-1 Mpc - intermediate LSS scales (galaxies + clusters);• supercomputing/multiprocessing SGI-ALTIX (44 proc.)• grid 5123/10243 with a mass resolution of ~105 M ;• initial conditions: periodical, linear cosmological, r(t,q)=a(t)[q-b(t)q(q)],

from Q era (Q perturbations) CMB deculping era (baryons+CDM, z=3000, t=64kyr from Big Bang).

Software: • discrete model: parallel computing Art-PM (P3M+SPH), • intensive calculations: 2 x O(NlogN).• identification + evolution of LSS filamentary structure;• identification + structure of halos galaxies (FOF codes; ellipticities,

angular momentum, alignments, radial profiles);• history and evolution of halos (FOF history, parental, descending);• specialized visualization code (TIPSY).

z=5

1.b Large Scale Structure of the Universe (LSS) simulations, AIRA HD fields cosmological evolutionHD fields cosmological evolution

from CMB era to now.from CMB era to now.

LSS simulations, AIRA

Z=3, t=2.2Gyr

LSS – hydrodynamical model

v(x,t)=v(x0,t)+grad(v)(x-x0) +O(|x-x0|)=

= vcm + (D+Ω)(x-x0) +O(|x-x0|)=

= vcm + D(x0,t)(x-x0) + ω x (x-x0) + O(|x-x0|)

G=grad(v)=D+ Ω, Dij, Ω ij=½ (∂vi/∂xj± ∂vj/∂ xi)

D –strain tensor, Ω –vortex ,

ω=(Ω32, Ω13, Ω21)=½rot(v) - vorticity ,

tr(D)=div(v) - expansion.

HD flows:

D=0 rigidity; div(v)=0 incompressible fluid;

Ω=0 non rotational flow; Ω>>D rotational flow.

Velocity field in PM simulation (z=11, t=0.42 Gyr)

Hydrodynamical flow: LSS Virgo zone , z=0

LSS hydrodynamical flow:[halos-squares, rot(v)-green, D-violet, Ω-red)]

D- voids (expansion), Ω-halos + filaments (contraction)

Hydrodynamics: D- (contraction)

HD alignments

• Study of axis alignment for inertial, velocity and tidal tensors:

[Dij]ext(H) [Iij]H [vCM, i]H ,

• aij=Dij-(Dii/3)δij

iij=Iij - (Iij/3) δij

[dij]<-->[iij]<-->[vCM,i].

• Time slices: z=11 (t=0.42 Gyr) – z=8 (t=0.65 Gyr).

Z=11

Z=10

Z=9

Z=8

1.c Global search heuristic, FOF, AIRAGalaxies and clusters of galaxies in

LSS

Global search heuristic, FOF, AIRA

Overview of FOF and visualization tool TIPSY:• FoF is a simple group finder, which uses the friends-of-friends

method to find groups. A particle belongs to a friends-of-friends group if it is within some linking length of any other particle in the group. After all such groups are found, those with less than a specified minimum number of group members are rejected. The program takes input files in the tipsy binary format and produces a single ASCII output file called fof.grp. This output file is in the tipsy array format and contains the group number to which each particle belongs. A group number of zero means that the particle does not belong to a group. The fof.grp file can be read in by tipsy and used to color each particle by group number. This provides an effective way to visualize the groups that are found by fof. Simulations with periodic boundary conditions can also be handled by fof by specifying the period in each dimension on the command line (see MAN page below).

• FOF intensive calculations: O(N2logN).

Global search heuristic, FOF, AIRA

Global search heuristic, FOF, AIRA

Global search heuristic, FOF, AIRA

Global search heuristic, FOF, AIRA

Global search heuristic, FOF, AIRA

Global search heuristic, FOF, AIRA

Global search heuristic, FOF, AIRA

HD Cosmological simulations, High z cosmological simulations resources

z=5

Grid Memory N-body Computer Speed Memory SPH

2563/5123 576 Mb SGI Pow. Ch. workstation 300 MFLOPS 1.1 Gb

5123/10243 4.6 Gb SGI ALTIX 350

4 proc. 12Gb

24 GFLOPS 9.2 Gb

10243/20483 36.8 Gb SGI ALTIX 3700

40 proc. 80Gb

240 GFLOPS 73.6 Gb

10243/20483 ~1 Tb IBM Power PC 970

512 proc.

Mare Nostrum simulation

Barcelona, Spain

4.5 TFLOPS ~1 Tb

(+ FOF)

20483/40963 294.4 Gb 128 proc.

512 proc. IBM p690

Millennium Simulation

UK, Germany, VIRGO

2.4 TFLOPS?

0.2 TFLOPS

2005

588 Gb

25 Tb total

(+FOF)

40963/81923 2355 Gb 1024 proc.

2048 proc.IBM Power6

Millennium II Simulation

UK, Germany, VIRGO

24 TFLOPS?

2009 ?

4710 Gb

?

(+FOF)

81923/163843 18840 Gb 4096 proc.

No

240 TFLOPS?

No

37680

No

Scaling factor memory (N-body): 8Scaling factor memory (N-body+SPH): 16Scaling factor speed: ~10z slices history (cosmological spacetime slices) ~ 100-1000

2.a.Machine Learning, AIRA

2.a.Machine Learning, AIRA Neural Network (Multilayer perceptron)

• N-body codes, phase space: [mi, (xi,yi,zi), (x’i, y’i, z’i)]

• Astronomy, coordinates 2D [αi, δi] (celestial sphere)

• The problem 2D 3D [αi, δi, zi], (α’i, δ’i, z’i)

z-coordonate (redshift) calculation

• Small distaces z estimated spectroscopic (zsp =Δλ/λ);

• Large distances z estimated photometric (zph = f(pi,wi))

• Relation zsp zph determined by a learning method;

• Learning method: Neural Network.

2.Machine Learning, AIRA Neural Network (Multilayer perceptron)

Data• HDFN: (Fernandez-Soto+, 1999) J/ApJ/513/34 [1067 galaxies]; • HDFS: HDF-South NICMOS field (Yahata+, 2000) J/ApJ/538/493 [335 galaxies];• CHANDRA: Chandra Deep Field South: multi-colour data catalog (Wolf+, 2004) [63500

galaxies];Colors: UBVIJHK ;Cosmological simulations:

– in HDF (UBVRIJKLMNH) + SDSS (UBVRI) – model LCDM :

• Ω0 Λ h Sigma8 Npar L(Mpc/h) mp(Msun/h) l_soft(Kpc/h)• 0.3 0.7 0.7 0.90 2563 239.5 6.86e10 25 Model LCDM+SES+SED:• fbar α ε fbulge MB,lum MB,morpho• 0.12 0.05 0.05 0.1 -16.25 -18.46Number of galaxies:• z 0 0.2 0.42 0.62 0.82 1.05 1.47 2.12 2.97• N 1178 938 731 612 519 439 341 232 156 Our Method : Neural Network (NN) reduced field HDFN+HDFS+ CHANDRA• Learning data: HDFN+HDFS_CHANDRA - 10000 galaxies with known zph si zsp (hyperz);• Validation data: 5000 galaxies with zsp available;• Test data– all data from catalogs HDFN,HDFS, CHANDRA.

2.Machine Learning, AIRA Neural Network (Multilayer perceptron)

zsp- zph in HDFN+HDFS. Red – our NN code, blue: HYPERZ NN-type code.

2.b. Global search heuristics, AIRA

2.b. Global search heuristics, AIRAGenetic algorithm – optimization problem

2.b. Global search heuristics, AIRA Stellar inversion method in asteroseismology

Investigation of stellar structures (M) using asteroseismological properties of stars (different observed pulsating periods (P)):

P Musing a minimization process (O-C)2.Quantitatively, we have:- N

- number of periods determined/star, degenerate in (n,l,m) modes;

- N - number of frequencies/filters/colors for observations;- NL – number of lines frequencies/line…;- NZ – number of elements others than X and Y ;- Nd – number of parameters for diffusion…..

• Consistent solutions if: Nr (P) >> Nr (M)

2.b. Global search heuristics, AIRA Stellar inversion method in asteroseismology(space missions KEPLER, CoRoT, MOST)

• Inverse methods (asteroseismological methods):

M(S) = min (L-1 [Pth- Pobs]) where we have:- Space of equations: P = [nlm, A,nlm, ,nlm, (lpv),k, nlm, nlm]

- Space of parameters: M = [(M, t0) {X, Y, ZjNz

}, MLT ,(i,vr,0) (DNd,….)] = = [(M, Te ) {X, Y, Zj

Nz }, MLT) , (DNd,….)]

- Space of solutions: S = [Xi, Yi] (stellar + pulsation)

Xi =[p, T, r, L, , , , CV,CP, 1, 3, Lrad, ad, rad, , T, T, , ,,d/dr]i

Yi =[(y1,…,y6) , , (≡I/R), (A,r,L,T,p)]i,nlm i= 1, N

2.b. Global search heuristics, AIRA Stellar inversion method in asteroseismology

• Our asteroseismological method:– CESAM (Cesam2k_V2)+ ROMOSC (linear, nonadiabatic, nonradial)

with an automated queue of calculations for entire tracks (pms-ms-postms) – Minimization of:

– For a grid of stellar evolutive tracks: [M, (X,Z), t ] [M, (X,Z), Te, L]

• Our calculations:– observed pulsating modes (g, p, l=0-3);– [R, I] [R, DZ] (theoretical excited and stable modes) 2 minima for a grid of tracks:

• M = 0.8 M - 6.0 M ;• Z = 0.01, 0.02, 0.03, 0.04;

– evolutive stages: from PMS (stellar formation) to AGB;• Calibration: COROT/MOST CESAM+ROMOSC/YREQ+QDG• Results: ~21.000.000 pulsation modes/250 days CPU, 1.5 Tb data

N

iiCiON 1

2,,

2 )(1

2.b. Global search heuristics, AIRA Romanian participation to KEPLER, CoRoT, MOST

space missions

Number of chemical compositions (Z) 4 Number of tracks (0.8M

-6M

, M=0.01) 50 Number of points (stars) per track (pms-postMS) 300 Number of pulsating modes per star (g,f,p, R[0.01, 50] ~300 Total number of pulsating modes ~20,000,000 Computing time/pulsating mode ~1s Total computing time 230 days (from August 2008) Storage/pulsating mode 0.1MB Total storage 2 TB Percent of proposed grid realized up to now 100%

•Theoretical calculations for grids of evolutionary tracks (0.8M-6M, Z=0.01-0.04, M=0.01-0.001M) using CESAM model + COEFNAD. (2007-2008).

•Theoretical calculations for grids of stellar pulsation modes using ROMOSC model (pre-postMS, B-Be, Delta Scuti Solar type stars, 0.8M-6M, g,f,p, Z=0.01-0.04, M=0.01-0.001M, Z=0.01).

•Based on these grids, theoretical calculations of stellar data using inverting methods (SEISMROM asteroseismological method) for the observed pulsating stars in KEPLER, CoRoT, MOST stace missions (precision 0.1Hz).

Needed theoretical grids:

2.b. Global search heuristics, AIRA GRID: Stellar evolutive traks, 1-6 Msun stars

2.b. Global search heuristics, AIRA GRID: Pulsating linear, non-adiabatic, non-radial

modes (some examples)

Romanian participation to KEPLER, CoRoT, MOST space missions

• Our results concern the following stars :

– Solar type stars: Sun (128 modes, used as calibrator), Boo (40 modes, MOST), 5500 stars (KEPLER);

Scuti/ Doradus /red giant stars: V2367 Cyg (368 modes,KEPLER), Oph (80 modes,MOST), BD+184914 (16 modes, KEPLER), K03219256 (16 modes, KEPLER), K07119530 (13 modes,KEPLER), K08583770 (9 modes,KEPLER), K572440 (9 modes, KEPLER), K7798339 (8 modes,KEPLER), K8355130 (8 modes, KEPLER);

– Ground observed pre/postMS/WVir stars: V351 Ori (5 modes/ post-MS, CoRoT);

– SPB stars: HD163830 (18 modes, KEPLER), K9716456 (7 modes, KEPLER), K11973705 (7 modes, KEPLER), K12258330 (13 modes, KEPLER), K7798339 (6 modes, KEPLER);

2.b. Global search heuristics, AIRA 5500 solar-like stars observed by KEPLER space mission

(red – theoretical grids, blue – theoretical results)

2.b. Global search heuristics, BLUE GENE + TERA GRID/US

Stellar inversion method in asteroseismology/adiabatic models MPIKAIA

Welcome to the TeraGrid Asteroseismic Modeling Portal

• The Asteroseismic Modeling Portal (AMP) provides a web-based interface for astronomers to use the Aarhus Stellar Evolution Code (ASTEC) coupled with a parallel genetic algorithm (MPIKAIA) to derive the properties of Sun-like stars from observations of their pulsation frequencies. For example, check out the data from our favorite star, the sun.

• What can I do with AMP?• Everyone can browse the catalog of runs to find data about stars that have been modeled

with AMP. You can find basic data about stars, such as their mass and metallicity, and download a Hertzsprung-Russell (HR) diagram that shows the star's temperature and luminosity during its lifetime.

• Scientists can use AMP to do two things: • Observable parameter optimization. In an optimization run, a scientist specifies

observable properties, such as pulsation frequencies, and a genetic algorithm is used to identify the stellar model that best fits the observed data. An optimization run makes extensive usage of BLUE GENE + TERAGRID computational resources and consumes about 20,000 CPU hours.

• Direct ASTEC model evaluation. In a direct model run, a scientist specifies a star's parameters, and the specific star is modeled using ASTEC. Data about the hypothetical star, including Echelle and HR diagrams, are produced for inspection. A direct model evaluation takes about 15 minutes to run.

• For more information about AMP, please see About AMP and the AMP User Guide. • AMP Supports Kepler Science• The Kepler satellite promises to revolutionize the quality and quantity of asteroseismic

data available for solar-type stars. The mission will use asteroseismology to determine precise absolute sizes of the potentially habitable Earth-like planets that are discovered. A uniform analysis of the asteroseismic data will help minimize systematic errors between the parameter values obtained for different stars, and such an analysis will be facilitated by this TeraGrid +BLUE GENE-based community modeling tool.

• AMP is a TeraGrid + BLUE GENE Science Gateway• The genetic algorithm that performs the asteroseismology science requires immense

amounts of computational time to run. AMP uses TeraGrid + BLUE GENE computational resources to perform its simulations, including systems located at the National Center for Atmospheric Research and the Texas Advanced Computing Center.

Stellar Astronomy resourcesModel Type Nr. Processors Calculus Model

Stellar structure

Stellar evolution

1 evolutionary track

4 eq. x 4 unkn.

1D

Nonlinear

1 proc./1h CPU REAL DUBLE PRECISION

HENTEY, CESAM2k_V2

HD

Stellar pulsation

Asteroseismology

1 pulsating track

adiabatic

4 eq. x 4 unkn.

3D

linear,nonradialnonadiabatic

(l,n,m)

Nonlinear system

1proc. /5h CPU REAL DUBLE PRECISION

SGI ALTIX /noi (ROMOSC)

BLUE GENE/US (MPIKAIA)

HD+rot

Stellar pulsation

Asteroseismology

1 pulsating track

Nonadiabatic

6 eq x 6 unkn.

3D

linear,nonradialnonadiabatic

(l,n,m)

Nonlinear system

1 proc. /250h CPU

1 pulsational mode/

1 proc./1 min CPU

1 pulsational mode/75kb

COMPLEX DUBLE PRECISION

SGI ALTIX/noi (ROMOSC)

64-bit (separation poles/roots)

32-bit BLUE GENE not enough rigorous mathematics

NO BLUE GENE ???

HD+rot

Stellar structure

1 stellar structure

3D

Nonlinear

turbulent flows

100 – 10000 proc.

10000 proc/300h CPU

(Moscow,Mil. Academy)

REAL DUBLE PRECISION

(in implementation Moscow model/ Gudunov Scheme)

MHD+rot+

+conv+turb

Solar model

Supernovae

3. Robotic/thinking and remote

control 1.3m Romanian

telescope, AIRA

3. Robotic/thinking telescope Science ProgrammeRobotic/thinking and remote control 1.3m Romanian telescope.

A robotic telescope is an astronomical telescope and detector system that makes observations without the intervention of a human. In astronomical disciplines, a telescope qualifies as robotic if it makes those observations without being operated by a human, even if a human has to initiate the observations at the beginning of the night, or end them in the morning. A robotic telescope is distinct from a remote telescope, though an instrument can be both robotic and remote.

• Robotic telescopes are complex systems that typically incorporate a number of subsystems. These subsystems include devices that provide telescope pointing capability, operation of the detector (typically a CCD camera), control of the dome or telescope enclosure, control over the telescope's focuser, detection of weather conditions, and other capabilities. Frequently these varying subsystems are presided over by a master control system, which is almost always a software component.

• Robotic telescopes operate under closed loop or open loop principles. – In an open loop system, a robotic telescope system points itself and collects its

data without inspecting the results of its operations to ensure it is operating properly. An open loop telescope is sometimes said to be operating on faith, in that if something goes wrong, there is no way for the control system to detect it and compensate.

– A closed loop system has the capability to evaluate its operations through redundant inputs to detect errors. A common such input would be position encoders on the telescope's axes of motion, or the capability of evaluating the system's images to ensure it was pointed at the correct field of view when they were exposed.

automaton/autonomous robot (thinking/self-operating system)

Robotic/thinking 1.3m telescope, AIRA

• In 2002, the RAPid Telescopes for Optical Response (RAPTOR) project pushed the envelope of automated robotic astronomy by becoming the first fully autonomous closed–loop robotic telescope. RAPTOR was designed in 2000 and began full deployment in 2002. Its first light on one of the wide field instruments was in late 2001, with the second wide field system coming online in early 2002. Closed loop operations began in 2002. Originally the goal of RAPTOR was to develop a system of ground-based telescopes that would reliably respond to satellite triggers and more importantly, identify transients in real-time and generate alerts with source locations to enable follow-up observations with other, larger, telescopes. It has achieved both of these goals quite successfully. Now RAPTOR has been re-tuned to be the key hardware element of the Thinking Telescopes Technologies Project.

• Its new mandate will be the monitoring of the night sky looking for interesting and anomalous behaviors in persistent sources using some of the most advanced robotic software ever deployed. The two wide field systems are a mosaic of CCD cameras. The mosaic covers and area of approximately 1500 square degrees to a depth of 12th magnitude. Centered in each wide field array is a single fovea system with a field of view of 4 degrees and depth of 16th magnitude. The wide field systems are separated by a 38km baseline. Supporting these wide field systems are two other operational telescopes. The first of these is a cataloging patrol instrument with a mosaic 16 square degree field of view down to 16 magnitude. The other system is a .4m OTA with a yielding a depth of 19-20th magnitude and a coverage of .35 degrees. Three additional systems are currently undergoing development and testing and deployment will be staged over the next two years. All of the systems are mounted on custom manufactured, fast-slewing mounts capable of reaching any point in the sky in 3 seconds. The RAPTOR System is located on site at Los Alamos National Laboratory (USA) and has been supported through the Laboratory's Directed Research and Development funds.

Robotic/thinking 1.3m telescope, AIRA

Thinking Telescope (new Romanian 1.3m telescope)Why a Thinking Telescope?• The existence of rapidly slewing robotic telescopes and fast alert distribution

via the Internet is revolutionizing our capability to study the physics of fast astrophysical transients. But the salient challenge that optical time domain surveys must conquer is mining the torrent of data to recognize important transients in a scene full of normal variations. Humans do not have the ability to recognize fast transients and rapidly respond. Autonomous robotic instrumentation with the ability to extract pertinent information from the data stream in real time will therefore be essential for recognizing transients and commanding rapid follow-up observations while the ephemeral behavior is still present.

• The development and integration of three technologies: (1) robotic telescope networks; (2) machine learning; and (3) advanced database technology, can enable the construction of smart robotic telescopes, which we loosely call ?thinking? telescopes, capable of mining the sky in real time.

• The Thinking Telescope is a concept designed to enhance analysis and observation of huge sections of the night sky and to be able to extract useful information in a timely fashion. 

Robotic/thinking 1.3m telescope, AIRA

Thinking telescope. Three technologies must be integrated in order to create autonomous robotic telescope systems capable of finding and making more detailed follow-up observations of ephemeral source anomalies in real time:

• Robotic hardware,• Machine learning,• Context knowledge.

Automate reasoning Artificial inteligence

Robotic/thinking 1.3m telescope, AIRA

To construct thinking telescopes for astronomy, we must develop and integrate technology in three key areas: • Distributed networks of robotic monitoring and response telescopes • Machine learning techniques for automated knowledge extraction in real time. • Virtual observatories employing advanced database technology that provide context information in real time

CONCLUSIONSSupercomputing, AIRA

Very different supercomputing directions, AIRA:

• Computational fluid dynamics: MHD HD Nbody+SPH.

• Calculations to understand the fundamental nature of matter Universe formation and evolution/from the inflation period to now.

• Artificial intelligence / Symbolic computations including: Recognition processes, Automated reasoning, Global search heuristics

END,

Thank you for your kind invitation!

AI supercomputingArtificial inteligence:Tools• In the course of 50 years of research, AI has developed a large number of tools to solve the most difficult

problems in computer science. A few of the most general of these methods are discussed below.

3. Neural networks• Main articles: Neural networks and Connectionism• A neural network is an interconnected group of nodes, akin to the vast network of neurons in the human brain.• The study of artificial neural networks[125] began in the decade before the field AI research was founded, in the

work of Walter Pitts and Warren McCullough. Other important early researchers were Frank Rosenblatt, who invented the perceptron and Paul Werbos who developed the backpropagation algorithm.[132]

• The main categories of networks are acyclic or feedforward neural networks (where the signal passes in only one direction) and recurrent neural networks (which allow feedback). Among the most popular feedforward networks are perceptrons, multi-layer perceptrons and radial basis networks.[133] Among recurrent networks, the most famous is the Hopfield net, a form of attractor network, which was first described by John Hopfield in 1982.[134] Neural networks can be applied to the problem of intelligent control (for robotics) or learning, using such techniques as Hebbian learning and competitive learning.[135]

• Jeff Hawkins argues that research in neural networks has stalled because it has failed to model the essential properties of the neocortex, and has suggested a model (Hierarchical Temporal Memory) that is loosely based on neurological research.[136]

4. Control theory• Main article: Intelligent control• Control theory, the grandchild of cybernetics, has many important applications, especially in robotics.[137]

AI supercomputingMachine learning• Machine learning is a scientific discipline that is concerned with the

design and development of algorithms that allow computers to evolve behaviors based on empirical data, such as from sensor data or databases. A major focus of machine learning research is to automatically learn to recognize complex patterns and make intelligent decisions based on data; the difficulty lies in the fact that the set of all possible behaviors given all possible inputs is too complex to describe generally in programming languages, so that in effect programs must automatically describe programs. Artificial intelligence is a closely related field, as also probability theory and statistics, data mining, pattern recognition, adaptive control, and theoretical computer science.

2.b. Global search heuristics, AIRA Stellar inversion method in asteroseismology

Pulsating stars:• Direct methods: Pth = K[M(S)] • Comparison with observations:

Pth- Pobs = L[M(S)] 0• Inverse methods (asteroseismological methods):

M(S) = min (L-1 [Pth- Pobs]) where we have:- Space of equations: P = [Pnlm, A,nlm, ,nlm, (lpv),k, nlm, nlm]

- Space of parameters: M = [(M, t0) {X, Y, ZjNz

}, lMLT ,(i,vr,0) (DNd,….)] = = [(M, Te, L ) {X, Y, Zj

Nz }, lMLT) , (DNd,….)]

- Space of solutions: S = [S1,i, S2,i] (stellar + pulsation)

S1,i =[p, T, r, L, , , , CV,CP, 1, 3, Lrad, ad, rad, , T, T, , ,,d/dr]i

S2,i =[(y1,…,y6) , , (≡I/R), r,L,T,p]i,nlm i= 1, N

AI supercomputingMachine learning Algorithm types

Machine learning algorithms are organized into a taxonomy, based on the desired outcome of the algorithm:

• Supervised learning generates a function that maps inputs to desired outputs. For example, in a classification problem, the learner approximates a function mapping a vector into classes by looking at input-output examples of the function.

• Unsupervised learning models a set of inputs, like clustering.

• Semi-supervised learning combines both labeled and unlabeled examples to generate an appropriate function or classifier.

• Reinforcement learning learns how to act given an observation of the world. Every action has some impact in the environment, and the environment provides feedback in the form of rewards that guides the learning algorithm.

• Transduction tries to predict new outputs based on training inputs, training outputs, and test inputs.

• Learning to learn learns its own inductive bias based on previous experience.

AI supercomputing

Supervised learning

• Supervised learning is a machine learning technique for deducing a function from training data. The training data consist of pairs of input objects (typically vectors), and desired outputs.

• Supervised learning can generate models of two types. Most commonly, supervised learning generates a global model that maps input objects to desired outputs. In some cases, however, the map is implemented as a set of local models (such as in case-based reasoning or the nearest neighbor algorithm).

AI supercomputingThe most widely used classifiers are:• Neural Network (Multilayer perceptron),

• Support Vector Machines,

• k-nearest neighbor algorithm,

• Gaussian mixture model,

• Gaussian, naive Bayes,

• decision tree,

• radial basis function classifiers.

AI supercomputingNeural networks• An artificial neural network (ANN), usually called "neural network" (NN), is a

mathematical model or computational model that tries to simulate the structure and/or functional aspects of biological neural networks. It consists of an interconnected group of artificial neurons and processes information using a connectionist approach to computation. In most cases an ANN is an adaptive system that changes its structure based on external or internal information that flows through the network during the learning phase. Modern neural networks are non-linear statistical data modeling tools. They are usually used to model complex relationships between inputs and outputs or to find patterns in data.

• Perhaps the greatest advantage of ANNs is their ability to be used as an arbitrary function approximation mechanism which 'learns' from observed data. However, using them is not so straightforward and a relatively good understanding of the underlying theory is essential.

• Choice of model: This will depend on the data representation and the application. Overly complex models tend to lead to problems with learning.

• Learning algorithm: There are numerous tradeoffs between learning algorithms. Almost any algorithm will work well with the correct hyperparameters for training on a particular fixed dataset. However selecting and tuning an algorithm for training on unseen data requires a significant amount of experimentation.

• Robustness: If the model, cost function and learning algorithm are selected appropriately the resulting ANN can be extremely robust.

• With the correct implementation ANNs can be used naturally in online learning and large dataset applications. Their simple implementation and the existence of mostly local dependencies exhibited in the structure allows for fast, parallel implementations in hardware.

AI supercomputingEvolutionary algorithm

In artificial intelligence, an evolutionary algorithm (EA) is a subset of evolutionary computation, a generic populationbased metaheuristic optimization algorithm. An EA uses some mechanisms inspired by biological evolution:

reproduction, mutation, recombination, and selection. Candidate solutions to the optimization problem play the role of individuals in a population, and the fitness function determines the environment within which the solutions "live" (see also cost function). Evolution of the population then takes place after the repeated application of the above operators. Artificial evolution (AE) describes a process involving individual evolutionary algorithms; EAs are individual components that participate in an AE.

Evolutionary algorithm techniquesSimilar techniques differ in the implementation details and the nature of the particular applied problem:• Genetic algorithm - This is the most popular type of EA. One seeks the solution of a problem in the form of strings

of numbers (traditionally binary, although the best representations are usually those that reflect something about the problem being solved), by applying operators such as recombination and mutation (sometimes one, sometimes both). This type of EA is often used in optimization problems;

• Genetic programming - Here the solutions are in the form of computer programs, and their fitness is determined by their ability to solve a computational problem.

• Evolutionary programming - Similar to genetic programming, but the structure of the program is fixed and its numerical parameters are allowed to evolve;

• Evolution strategy - Works with vectors of real numbers as representations of solutions, and typically uses self-adaptive mutation rates;

• Neuroevolution - Similar to genetic programming but the genomes represent artificial neural networks by describing structure and connection weights. The genome encoding can be direct or indirect.

AI supercomputingGenetic algorithms• A genetic algorithm (GA) is a search technique used in computing

to find exact or approximate solutions to optimization and search problems. Genetic algorithms are categorized as global search heuristics. Genetic algorithms are a particular class of evolutionary algorithms (EA) that use techniques inspired by evolutionary biology such as inheritance, mutation, selection, and crossover.

• Genetic algorithms are implemented in a computer simulation in which a population of abstract representations (called chromosomes or the genotype of the genome) of candidate solutions (called individuals, creatures, or phenotypes) to an optimization problem evolves toward better solutions.

AI supercomputingNeural networks

The tasks to which artificial neural networks are applied tend to fall within the following broad categories:

• Global search Heuristics,

• Function approximation, or regression analysis, including time series prediction, fitness approximation and modeling.

• Classification, including pattern and sequence recognition, novelty detection and sequential decision making.

• Data processing, including filtering, clustering, blind source separation and compression.

• Robotics, including directing manipulators, Computer numerical control.