OpenAI-5 vinder over Dota2 amatører

Max Tegmark: Intelligence = ability to accomplish complex goals

Man anvender i dette tilfælde “reinforcement learning” ved at lade fem neurale netværk spille en eSport mod hinanden ved anvendelse af en optimeringsmetode, som man kalder “Proximal Policy Optimization”. OpenAI Five træner altså ikke ved at spille mod mennesker.

OpenAI Five

Our team of five neural networks, OpenAI Five, has started to defeat amateur human teams at Dota 2. While today we play with restrictions, we aim to beat a team of top professionals at The International in August subject only to a limited set of heroes. We may not succeed: Dota 2 is one of the most popular and complex esports games in the world, with creative and motivated professionals who train year-round to earn part of Dota’s annual $40M prize pool (the largest of any esports game).

OpenAI Five plays 180 years worth of games against itself every day, learning via self-play. It trains using a scaled-up version of Proximal Policy Optimization running on 256 GPUs and 128,000 CPU cores — a larger-scale version of the system we built to play the much-simpler solo variant of the game last year. Using a separate LSTM for each hero and no human data, it learns recognizable strategies. This indicates that reinforcement learning can yield long-term planning with large but achievable scale — without fundamental advances, contrary to our own expectations upon starting the project.

Follow the links to learn more.

Robotter er klar til at vinde VM i kendt computerspil

Robotterne har været i træningslejr. I meget lang tid. Faktisk i det, der svarer til at spille hver dag i 180 år.

Så lang tid har et hold nye robotter trænet for at blive klar til verdensmesterskaberne i computerspillet Dota 2. Turneringen er en af verdens største årlige begivenheder inden for lige præcis det spil.

Mesterskabet, der hedder The International, finder sted i august, og som noget nyt får verdens bedste menneskelige computerspillere nu konkurrence af et hold bestående af bots med kunstig intelligens.

Den kunstige intelligens er udviklet af tech-virksomheden OpenAI, som blandt andre Elon Musk var med til at stifte i 2015.

 

Non-gravitational acceleration of ‘Oumuamua

Non-gravitational acceleration in the trajectory of 1I/2017 U1 (‘Oumuamua)

‘Oumuamua (1I/2017 U1) is the first known object of interstellar origin to have entered the Solar System on an unbound and hyperbolic trajectory with respect to the Sun. Various physical observations collected during its visit to the Solar System showed that it has an unusually elongated shape and a tumbling rotation state and that the physical properties of its surface resemble those of cometary nuclei, even though it showed no evidence of cometary activity. The motion of all celestial bodies is governed mostly by gravity, but the trajectories of comets can also be affected by non-gravitational forces due to cometary outgassing. Because non-gravitational accelerations are at least three to four orders of magnitude weaker than gravitational acceleration, the detection of any deviation from a purely gravity-driven trajectory requires high-quality astrometry over a long arc. As a result, non-gravitational effects have been measured on only a limited subset of the small-body population. Here we report the detection, at 30σ significance, of non-gravitational acceleration in the motion of ‘Oumuamua. We analyse imaging data from extensive observations by ground-based and orbiting facilities. This analysis rules out systematic biases and shows that all astrometric data can be described once a non-gravitational component representing a heliocentric radial acceleration proportional to r−2 or r−1 (where r is the heliocentric distance) is included in the model. After ruling out solar-radiation pressure, drag- and friction-like forces, interaction with solar wind for a highly magnetized object, and geometric effects originating from ‘Oumuamua potentially being composed of several spatially separated bodies or having a pronounced offset between its photocentre and centre of mass, we find comet-like outgassing to be a physically viable explanation, provided that ‘Oumuamua has thermal properties similar to comets.

 

KIC 8462852: Variable Secular Decline

The KIC 8462852 Light Curve From 2015.75 to 2018.18 Shows a Variable Secular Decline

The star KIC 8462852 (Boyajian’s Star) displays both fast dips of up to 20% on time scales of days, plus long-term secular fading by up to 19% on time scales from a year to a century. We report on CCD photometry of KIC 8462852 from 2015.75 to 2018.18, with 19,176 images making for 1,866 nightly magnitudes in BVRI. Our light curves show a continuing secular decline (by 0.023 +- 0.003 mags in the B-band) with three superposed dips with duration 120-180 days. This demonstrates that there is a continuum of dip durations from a day to a century, so that the secular fading is seen to be by the same physical mechanism as the short-duration Kepler dips. The BVRI light curves all have the same shape, with the slopes and amplitudes for VRI being systematically smaller than in the B-band by factors of 0.77 +- 0.05, 0.50 +- 0.05, and 0.31 +- 0.05. We rule out any hypothesis involving occultation of the primary star by any star, planet, solid body, or optically thick cloud. But these ratios are the same as that expected for ordinary extinction by dust clouds. This chromatic extinction implies dust particle sizes going down to ~0.1 micron, suggesting that this dust will be rapidly blown away by stellar radiation pressure, so the dust clouds must have formed within months. The modern infrared observations were taken at a time when there was at least 12.4% +- 1.3% dust coverage (as part of the secular dimming), and this is consistent with dimming originating in circumstellar dust.

 

21cm absorption from Marion

Probing Radio Intensity at high-Z from Marion: 2017 Instrument

We introduce Probing Radio Intensity at high-Z from Marion (PRIZM), a new experiment designed to measure the globally averaged sky brightness, including the expected redshifted 21 cm neutral hydrogen absorption feature arising from the formation of the first stars. PRIZM consists of two dual-polarization antennas operating at central frequencies of 70 and 100 MHz, and the experiment is located on Marion Island in the sub-Antarctic. We describe the initial design and configuration of the PRIZM instrument that was installed in 2017, and we present preliminary data that demonstrate that Marion Island offers an exceptionally clean observing environment, with essentially no visible contamination within the FM band.

 

Dipping events of KIC 8462852

The Variable Wavelength Dependence of the Dipping event of KIC 8462852

First observed with the Kepler mission, KIC 8462852 undergoes unexplained dimming events, “dips,” on the timescale of days which were again observed from the ground from May to December 2017. Monitored with multi-band photometry by the Los Cumbres Observatory, all four dips of the “Elsie dip family” display clear wavelength dependence. We measure how the wavelength dependence changes over the whole dimming event, including the dimming between the dips and the brightening event (the `blip’) which occurs after the dips. We find that a single wavelength dependence does not fit the entire light curve and the dimming occurring between the dips is non-gray and varies in time. Because of the non-gray dimming between the dips, we measure the wavelength dependence of the dips separately and without the extra depth from this dimming. Such measurements yield a different estimate of the wavelength dependence the wavelength dependence of the dips but remains consistent with the previous measurement except for Elsie (the first dip), which is surrounded by dimming with strong wavelength dependence. We find the range of the wavelength dependence variation of the entire 2017 light curve is consistent with optically-thin dust with an average radius of r < 1 μm and the dust causing just the dips being r < 0.5 μm. Since the dependence is time-dependent, the dust occulting the star must be heterogeneous in size, composition, or both and the distributions of these properties along the line of sight must change over time.

 

Two Twin Type Ia Supernovae

Significant Luminosity Differences of Two Twin Type Ia Supernovae

The Type Ia supernovae (SNe Ia) 2011by, hosted in NGC 3972, and 2011fe, hosted in M101, are optical “twins,” having almost identical optical light-curve shapes, colours, and near-maximum-brightness spectra. However, SN 2011fe had significantly more ultraviolet (UV; 1600 < λ < 2500 A) flux than SN 2011by before and at peak luminosity. Theory suggests that SNe Ia with higher progenitor metallicity should (1) have additional UV opacity near peak and thus lower UV flux; (2) have an essentially unchanged optical spectral-energy distribution; (3) have a similar optical light-curve shape; and (4) because of the excess neutrons, produce more stable Fe-group elements at the expense of radioactive 56Ni and thus have a lower peak luminosity. Foley & Kirshner (2013) suggested that the difference in UV flux between SNe 2011by and 2011fe was the result of their progenitors having significantly different metallicities. The SNe also had a large, but insignificant, difference between their peak absolute magnitudes (ΔMV, peak = 0.60 ± 0.36 mag), with SN 2011fe being more luminous. We present a new Cepheid-based distance to NGC 3972, significantly improving the precision of the distance measurement for SN 2011by. With these new data, we determine that the SNe have significantly different peak luminosities (ΔMV, peak = 0.335 ± 0.069 mag), corresponding to SN 2011fe having produced 38% more 56Ni than SN 2011by, and providing additional evidence for progenitor metallicity differences for these SNe. We discuss how progenitor metallicity differences can contribute to the intrinsic scatter for light-curve-shape-corrected SN luminosities, the use of “twin” SNe for measuring distances, and implications for using SNe Ia for constraining cosmological parameters.

 

Sandsynligheden for abiogenesis

On the Rate of Abiogenesis from a Bayesian Informatics Perspective

Life appears to have emerged relatively quickly on the Earth, a fact sometimes used to justify a high rate of spontaneous abiogenesis (λ) among Earth-like worlds. Conditioned upon a single datum – the time of earliest evidence for life (tobs) – previous Bayesian formalisms for the posterior distribution of λ have demonstrated how inferences are highly sensitive to the priors. Rather than attempt to infer the true λ posterior, we here compute the relative change to λ when new experimental/observational evidence is introduced. By simulating posterior distributions and resulting entropic information gains, we compare three experimental pressures on λ: 1) evidence for an earlier start to life, tobs, 2) constraints on spontaneous abiogenesis from the lab, and 3) an exoplanet survey for biosignatures. First, we find that experiments 1) and 2) can only yield lower limits on λ, unlike 3). Second, evidence for an earlier start to life can yield negligible information on λ if tobs ≪ 1/λmax. Vice versa, experiment 2) is uninformative when λmax ≫ 1/tobs. Whilst experiment 3) appears the most direct means of measuring λ, we highlight that early starts inform us of the conditions of abiogenesis, and that lab experiments could succeed in building new life. Together then, the three experiments are complimentary and we encourage activity in all to solve this grand challenge.

The probability distribution for the number of abiogenesis events that actually occur is described by a Poisson distribution P(N|λ,t). The probability of N abiogenesis events having transpiret during a time interval t is given by

P(N|λ,t) = exp(-λt)(λt)N/N!

P(N > 0|λ,t) = 1 – P(N = 0|λ,t) = 1 – exp(-λt)

 

CubeSats for Gravitational Wave Detection

SAGE: using CubeSats for Gravitational Wave Detection

SAGE (SagnAc interferometer for Gravitational wavE) is a fast track project for a space observatory based on multiple 12-U CubeSats in geostationary orbit. The objective of this project is to create a Sagnac interferometer with 73000 km circular arms. The geometry of the interferometer makes it especially sensitive to circularly polarized gravitational waves at frequency close to 1 Hz. The nature of the Sagnac measurement makes it almost insensitive to position error, allowing spacecrafts in ballistic trajectory. The light source and recombination units of the interferometer are based on compact fibered technologies, without the need of an optical bench. The main limitation would come from non-gravitational acceleration of the spacecraft. However, conditionally upon our ability to post-process the effect of solar wind, solar pressure and thermal expansion, we would detect gravitational waves with strains down to 10-21 over a few days of observation.

 

Fotometriske rødforskydninger fra AI

Flere panoramiske kortlægninger af fjerne galakser er under forberedelse (LSST, Euclid, WFIRST). Disse vil producere flerbåndsfotometri for milliarder af galakser, hvortil pålidelige rødforskydninger også kræves til studiet af universets storskalastruktur. Spektroskopiske rødforskydninger er imidlertid ekstremt tidskrævende, hvorfor det bliver nødvendigt at anvende fotometriske rødforskydninger. Den sande middelrødforskydning for objekter i hvert fotometrisk rødforskydningsbin må kendes til bedre end ∼0.002(1+z) (z er rødforskydningen) med strenge krav til brøkdelen af katastrofalt afvigende værdier. En yderligere udfordring er bestemmelsen af en robust sandsynlighedsfordeling for rødforskydningen, samt en fuldstændig forståelse af usikkerheder forbundet med enhver kosmologisk måling.

Man har traditionelt anvendt to metodet: skabelontilpasning og maskinlæringsalgoritmer (kunstig intelligens, dvs AI). En begrænsning ved begge metoder har hidtil været de fotometriske målinger, som er påvirkede af den valgte blændestørrelse, PSF-variationer og overlappende objekter. Vi har i de senere år set en revolution inden for anvendelsen af Deep Learning (dvs kunstig intelligens) til billedklassifikation. B. Hoyle har vist, at Deep Convolutional neural network (CNN) var i stand til at bestemme nøjagtige fotometriske rødforskydninger direkte ud fra multifarvebilleder.

Denne artikel beskriver udviklingen af en kunstig intelligens i form af et Deep Convolutional Neural Network til estimering af fotometriske rødforskydninger med tilhørende sandsynlighedsfordelinger for et udvalg af galakser fra the Sloan Digital Sky Survey med rødforskydninger z < 0.4. Metoden anvender al informationen i billederne. Som input anvendes 64×64 pixler for hver af de fem optiske filtre ugriz centreret omkring de spektroskopiske mål, plus Mælkevejens rødfarvning langs retningen til disse mål. Dette CNN trænes på mere end 100000 objekter med spektroskopiske rødforskydninger, hvorefter det er i stand til at bestemme fotometriske rødforskydninger med en Mean Absolute Deviation (MAD) < 0.01, hvilket er signifikant lavere end de bedste spredninger for alternative metoder anvendt på de samme data. Den systematiske fejl er mindre end 0.0001, uafhængig af rødforskydningen. Forfatterne finder desuden, at CNN-rødforskydningen er uafhængige af en galakses hældning. Spredningen aftager med signal/støj-forholdet (SNR), og opnår værdier under 0.007 for SNR > 100. Disse resultater med kunstig intelligens lover godt for de planlagte kortlægninger af universets storskalastruktur.

Photometric redshifts from SDSS images using a Convolutional Neural Network

We developed a Deep Convolutional Neural Network (CNN), used as a classifier, to estimate photometric redshifts and associated probability distribution functions (PDF) for galaxies in the Main Galaxy Sample of the Sloan Digital Sky Survey at z < 0.4. Our method exploits all the information present in the images without any feature extraction. The input data consist of 64×64 pixel ugriz images centered on the spectroscopic targets, plus the galactic reddening value on the line-of-sight. For training sets of 100k objects or more (≥ 20% of the database), we reach a dispersion σMAD<0.01, significantly lower than the current best one obtained from another machine learning technique on the same sample. The bias is lower than 0.0001, independent of photometric redshift. The PDFs are shown to have very good predictive power. We also find that the CNN redshifts are unbiased with respect to galaxy inclination, and that σMAD decreases with the signal-to-noise ratio (SNR), achieving values below 0.007 for SNR >100, as in the deep stacked region of Stripe 82. We argue that for most galaxies the precision is limited by the SNR of SDSS images rather than by the method. The success of this experiment at low redshift opens promising perspectives for upcoming surveys.

 

Da Gaia-Enceladus blev indfanget af Mælkevejen

Mælkevejens forskellige bestanddele med tilhørende rumlig fordeling, kinematik, alder og grundstofmæssig fordeling kan anvendes til at afsløre rækkefølgen af de begivenheder, som førte til dannelsen af Mælkevejen. Dette skyldes, at gamle stjerner med masser under Solens masse er næsten lige så gamle som Universet selv, og deres atmosfæresammensætning i stor udstrækning afspejler sammensætningen af gassen, hvori de blev dannet. Stjernernes baner indeholder desuden information om, hvor de blev dannet, og de kan således afsløre, om stjernerne blev dannet i en ekstern galakse, som blev indfanget og smeltede sammen med Mælkevejen.

Kosmologiske modeller forudsiger, at galakser som Mælkevejen vokser via indfangning af og sammensmeltning med mindre galaksesystemer. Eftersøgning efter aftryk fra indfangningen af mindre galakser i Mælkevejen startede for 20 år siden. Det har været en længerevarende gåde, om Mælkevejen har haft en usædvanlig rolig indfangningshistorie, som muligvis kunne være i modstrid med kosmologiens standardmodel.

The formation of two of the major structural components of the Milky Way

One of the main goals of modern astrophysics is to understand how galaxies form and evolve from the Big Bang until the present-time. The Gaia mission was conceived to unravel the assembly history of our own Galaxy, the Milky Way. Gaia’s recently delivered second data release allows tackling this objective like never before. Here we analyse the kinematics, chemistry, age and spatial distribution of stars in a relatively large volume around the Sun that are mainly linked to two major Galactic components, the thick disk and the stellar halo. We demonstrate that the inner halo is dominated by debris from an object slightly more massive than the Small Magellanic Cloud, and which we refer to as Gaia-Enceladus. The accretion of Gaia-Enceladus must have led to the dynamical heating of the precursor of the thick disk and hence contributed to the formation of this component approximately 10 Gyr ago.

Forfatterne anvender resultater fra Gaia-missionen i kombination med andre data til at konkludere, at Mælkevejens indre stellare halo er domineret af resterne fra en galakse, som var mere massiv end den lille magellanske sky. Denne galakse, som de kalder Gaia-Enceladus, faldt ind Mælkevejen for 10 milliarder år siden. Den relativt massive galakse medførte en dynamisk “opvarmning” af den eksisterende forløber til den tynde skive.  Stjerner, som dannes i en skive af gasskyer, starter med relativt lave tilfældige hastigheder omkring en cirkelbevægelse. Man siger, at stjernerne er dynamisk “kolde”. De relativt små tilfældige hastigheder vinkelret på skiven betyder, at skiven bliver tynd (på samme måde som Saturns ringe).  Hvis et tungt legeme falder gennem en tynd skive af stjerner, vil stjernernes tilfældige hastigheder omkring en cirkelbevægelse forøges drastisk. Man siger, at stjernerne “opvarmes” dynamisk. Der er ikke tale om en gastemperatur, men om stjernernes tilfældige hastigheder. De større tilfældige hastigheder vinkelret på skiven medfører, at skivens tykkelse forøges betydeligt. Man har længe vidst, at Mælkevejen foruden en tynd skive også har en tyk skive. Man ved nu, at den tykke skive er dannet ved indfangning af Gaia-Enceladus for 10 milliarder år siden.