Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	The four-epoch MASIV survey (Lovell et al. 2003, 2008) detected variability in 56% of the 482 sources ob- served but the characteristics of the variability changed from epoch to epoch. While able to illustrate this ISS intermittency, the MASIV survey itself, because of its limited duration and cadence, could not address many of the resulting questions, notably the extent to which the intermittency results from changes in the In order ISM properties or changes in the sources. to address the cause of the intermittency, a survey must have the following characteristics. It must ob- serve a large number of sources widely distributed on the sky so as to have robust statistics and to be able to search for correlations expected for ISS (e.g., a Galac- tic latitude dependence or with nearby structures in the ISM). The survey must measure both the total and linear polarization intensity of the sources—ISM- driven intermittency should result in the linear polar- ization scintillation tracking that of the total inten- sity, while source-driven intermittency should result in the linear polarization and total intensity scintillating nearly independently of each other. The survey must be conducted with close to daily cadence to capture the scintillation variations, and over a multi-year time scale in order to separate changes in source structure or scattering medium from changes associated with the annual cycle resulting from the Earth’s orbital velocity. In addition to determining the cause of ISS inter- mittency and mapping the scattering properties of the local ISM, such a survey would address the existence of AGN emission above the inverse Compton brightness limit and potentially constrain the baryonic content of the IGM as a function of redshift.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	complication arises because annual cycles in the vari- ability timescale (caused by the Earth’s velocity vec- tor changing with respect to the scattering screens) need to be distinguished from true IDV intermittency. However, these two phenomena can be separated by making regular (daily) observations over the course of at least two years. The variations should be dissimilar between epochs separated by a year if the scintillation is intermittent. Intermittency, if caused by the ISM, points to the physical nature of the turbulent patches responsible for IDV.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	AGN emission above the inverse Compton bright- ness limit: In some cases fast scintillation may be at- tributable to the appearance of micro-arcsecond struc- ture within a source, associated with an outburst (Macquart & de Bruyn 2007). The rise and fade times of the manifestation of fast variations allow us to determine the longevity of micro-arcsecond components associated with IDV: The rate at which the AGN supplies energy to power the bright emission that causes the variability, and the rate at which energy losses cause its eventual decay. These rates can then be compared with speciﬁc AGN processes to identify the mechanism associated with the super-Compton radio emission.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Baryonic content of the intergalactic medium: The MASIV survey showed that few sources above z ∼ 2 are variable (Lovell et al. 2008). If this appar- ent suppression of ISS as a function of source redshift is conﬁrmed by VAST, there are a number of possible explanations: (1) intrinsic source evolution; (2) gravi- tational microlensing; or (3) scattering in the turbulent ionised IGM. If the last is at least partially responsible, then the angular broadening is caused by the cumula- tive eﬀect of all the baryons in the ionised IGM. The sheer number of sources VAST can monitor will enable us to probe in detail the evolution of structure in the ionised IGM as a function of redshift.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	One potential complication with scintillation stud- ies is that the largest amplitude IDV is typically ob- served at frequencies near 5 GHz. At ASKAP frequen- cies, it is possible that the strength of scattering, even toward high Galactic latitudes, and intrinsic source diameters will result in large-amplitude scintillations being quenched (Dennison & Condon 1981), although lower amplitude refractive scintillations may still be exhibited (e.g., Dennison et al. 1987).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	power, and the accretion power (Merloni et al. 2003; Falcke et al. 2004) in active galactic nuclei. This re- lation can be extended to lighter black holes where it matches the correlation between the X-ray and ra- dio luminosities in XRB black hole candidates (BHCs) that was found by Gallo et al. (2003). Radio emission is also observed from XRBs containing weakly mag- netised neutron stars, such as Z and atoll sources, al- though they are at least one order of magnitude fainter in radio than the BHCs (e.g. Fender 2006), and a weak radio jet has also been observed from the cata- clysmic variable SS Cyg during a dwarf nova outburst (K¨ording et al. 2008). There is still not a good theo- retical understanding of how accretion discs produce outﬂows, but most models assume that the jet is ac- celerated by a magnetic ﬁeld (e.g. Blandford & Payne 1982). VAST will allow us to characterise the full range of transient jet phenomena across the fundamen- tal plane.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Magnetic ﬁelds can also act in other ways. They can convert the rotational energy of their source into electromagnetic energy and the mechanical energy of an outﬂow, as happens for instance in a radio pulsar, or they can release their own energy in the form of ﬂares on a star or the radio emission from Jupiter and (pos- sibly) Jupiter-like extrasolar planets. With VAST we will be able to reveal the unifying physical principles that underlie such diverse behaviour.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	As discussed in Section 2.2.3, Fiedler et al. (1987a) ob- served a sample of 33 AGN, with ﬂux densities of a few Janskys, every day for more than 6 years. All sources were variable over this time span with typi- cally smooth variations in ﬂux with quasi-periods of months to years. VAST will produce light curves and detect this sort of variability for more than 104 sources over the whole sky. The Fermi gamma-ray satellite ob- serves the whole sky every 3 hours and has detected blazar variability on timescales as short as 6 hours (Abdo et al. 2010). Simultaneous measurements of the gamma-ray and radio variability in blazars will provide us with a powerful tool for understanding the physics of the central engines in AGN.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	sion. Combining the radio observations with observa- tions at other wavelengths or even with gravitational waves would signiﬁcantly improve our understanding of galaxy mergers (e.g., Sesana et al. 2011).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Tidal disruption events (TDEs) occur when a star wan- ders too close to a black hole at the centre of a galaxy or star cluster. The black hole tidal forces tear the star apart, but the debris continue on ballistic orbits around the star and are accreted as they approach pe- riastron on the following orbit. The resulting sudden increase in the black hole accretion rate may be di- rectly detected at optical and X-ray frequencies. After the initial ﬂare up the luminosity decays as t−5/3 (Rees 1990). Several candidates of such TDEs have been de- tected serendipitously through all-sky X-ray and ultra- violet surveys (Grupe et al. 1999; Komossa & Greiner 1999; Komossa & Bade 1999; Esquej et al. 2007, 2008; Gezari et al. 2006, 2008, 2009; Lin et al. 2011; Gezari et al. 2012; Cenko et al. 2012a; Saxton et al. 2012).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Crucially, black hole accretion is often accompa- nied by relativistic jets, with a mechanical power often comparable to the radiative luminosity. The jet can be a source of beamed radio emission and hard X-rays in a way similar to a blazar. To a hard X-ray telescope such as the Burst Alert Telescope (BAT) on the Swift it will initially resemble a GRB, but it will be distin- guished by its much longer duration. Swift J1644+57 is the ﬁrst observed example of such a TDE candi- date, at redshift z = 0.35. It was discovered as a lu- minous X-ray transient and was observed across the electromagnetic spectrum, enabling multi-band anal- ysis and modeling of the event (Bloom et al. 2011; Levan et al. 2011; Zauderer et al. 2011; Berger et al. 2012; Wiersema et al. 2012). The mass of the cen- tral black hole responsible for this event was derived to be . 107M⊙ and the ejecta Lorentz factor & 10. More recently another TDE candidate has also been detected by Swift in hard X-rays, Swift J2058.4+0516 (Cenko et al. 2012b).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Radio afterglows of TDEs would be a powerful tool to understand jet physics and source geometry, through measurement of the ﬂux evolution and po- larization. The expected degree of polarization for these events is low if they are collimated, similar to GRB afterglows. Recent infra-red and radio polariza- tion measurements have been found to be consistent with theoretical expectations (Wiersema et al. 2012; Metzger et al. 2012).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Radio detections of TDE jets hold the key to other unsolved astrophysical problems (e.g. Krolik & Piran It is well known that not all accreting black 2012). holes have jets, but it is not clear why. Jets are formed only at certain ranges of accretion rate, but other pa- rameters such as the black hole spin may also be im- portant. Finding what fraction and what type of X-ray detected TDEs have radio jets will help address this question, and will shed light on the evolution of nuclear black holes in galaxies. Furthermore, the timescale for the formation of a radio-bright TDE jet will help us understand how quickly a collimated magnetic ﬁeld is generated inside the transient accretion disk.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	The predicted brightness of TDEs at gigahertz fre- quencies suggests that VAST should be able to detect about a dozen events per year. A systematic survey of the radio sky would be able to constrain the rate of TDEs and reduce the collimation uncertainty signiﬁ- cantly. In particular, the discovery of an entirely new population of orphan TDEs would allow us to place solid constraints on the occurrence rate of this phe- nomenon and therefore test the theoretical prediction that we should see ∼10−4 yr−1 per galaxy (Rees 1990).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Exploring the connection between accretion disks and jets in X-ray binaries (XRBs) requires correlated ra- dio and X-ray observations (Fender 2006). The VAST survey will reveal and subsequently measure the radio emission from a large number of XRBs and allow us to search for new radio outbursts. This, coupled with the recent availability of large ﬁeld of view instruments like Swift-BAT and Fermi-GBM, and with the early results (from 2013) of eROSITA’s all sky survey, will provide a large sample of XRBs with active jets, and will al- low us to do statistical studies of their time variability properties and X-ray/radio correlation.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	in the radiative power). We will also investigate what physical mechanism causes major radio ﬂares and the ejection of highly relativistic, optically-thin bullets, of- ten associated to the state transition from the hard/non- thermal to soft/thermal state.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	In addition to stellar mass and supermassive black holes, it is predicted that VAST will detect jets from intermediate mass black holes (IMBHs) with masses between ∼100 – 100,000 M⊙. IMBHs may provide a pathway for the formation of supermassive black holes (e.g. Ebisuzaki et al. 2001) and also have impor- tant connotations for other areas of astrophysics (e.g. Fornasa & Bertone 2008; Wang et al. 2010), but until recently their existence was highly disputed. The dis- covery of the object known as HLX-1 located in the galaxy ESO 243-49, with an X-ray luminosity ∼1000 times greater than the Eddington limit for a 10 M⊙ black hole, currently provides very strong evidence for the existence of IMBHs (Farrell et al. 2009), with in- dependent studies constraining the mass to between 9,000 M⊙ < MBH < 90,000 M⊙ (Davis et al. 2011; Servillat et al. 2011; Godet et al. 2012; Webb et al. 2012). Recently, transient radio emission was detected from HLX-1 following a transition from the low/hard to high/soft spectral states, consistent with the ejection of relativistic jets from an IMBH (Webb et al. 2012).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Objects such as as HLX1 may be very rare (only one known within 100 Mpc), but there are hundreds of ultraluminous X-ray sources (ULXs; see Feng & Soria 2011 for a recent review) in the local Universe with luminosities exceeding the classical Eddington limit of stellar-mass black holes. Most of them are likely to be explained through other mechanisms such as anisotropic emission and/or super-Eddington accretion, but we cannot rule out that a few of them may be IMBHs. In addition, there may be hundreds of low- state or quiescent IMBHs lurking in the core of glob- ular clusters, in the nuclei of dwarf galaxies, or inside recently merged galaxies that may occasionally become active. The detection of jet emission from these objects would allow us to infer the black hole mass, as the ra- dio and X-ray luminosities are correlated as a function of mass (e.g. Fender 2006). Steady jets would be unde- tectable for all but the closest of these objects. How- ever, transient radio ﬂares have been observed from XRBs associated with jet ejection events with radio lu- minosities a factor of 10 – 100 times brighter than the non-ﬂaring radio emission (e.g. K¨ording et al. 2006). Such ﬂares should be detectable by VAST out to dis- tances of ∼5 – 15 Mpc from 1,000 M⊙ black holes, and out to ∼23 – 72 Mpc from 10,000 M⊙ black holes (Far- rell, in preparation). Thus we will be able to survey the local Universe for ﬂaring IMBHs, placing constraints on the number density and therefore testing their va- lidity as a pathway for supermassive black hole forma- tion. Conversely, if ULXs are super-Eddington accre- tors, detection of radio emission will prove that even at those extreme accretion rates, a substantial fraction of accretion power can come out as kinetic power, and will improve our understanding of black hole feedback at high accretion rates.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	The merger of two compact objects (neutron stars) is expected to generate strong gravitational waves (§2.1.3) but they will also reveal themselves through electro- magnetic radiation. The relativistic, beamed outﬂows detected from short-hard GRBs are thought to origi- nate from such mergers (e.g Fong et al. 2012). VAST is not expected to fortuitously observe such beamed events. However, merger simulations also show that substantial mass can be ejected from these systems quasi-isotropically and at sub- or trans-relativistic ve- locities, and with energies of 1049−1050 erg (Piran et al. 2012).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Consider a mildly relativistic blast wave (β0 ∼ 0.8) moving out with an energy ∼ 1049 erg into the sur- rounding medium. The shock-wave sweeps the sur- rounding medium and heats it to relativistic tempera- tures by converting bulk kinetic energy of the incom- ing material into thermal energy of the shocked ma- terial. The electrons in the shock-heated plasma are accelerated in the post-shock magnetic ﬁeld to radi- ate synchrotron radiation. This would eventually be optically thin in the gigahertz regime, peaking on a timescale of about a couple of months when the blast wave starts to decelerate. The optimum waveband for detecting these transients, which are also supposed to be prime sources of gravitational waves, is around gi- gahertz frequencies (Nakar & Piran 2011). A merger within 300 Mpc would shine as a radio transient of about 1 mJy brightness at ∼1 GHz (Kamble & Ka- plan in preparation) which would be detectable to the VAST and, potentially, also to Advanced-LIGO. Both the peak ﬂux density and the time to peak are un- certain since we lack accurate estimates of the ejecta velocity and energy, and we expect the circumburst density to vary from 1 cm−3 to 10−4 cm−3.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Using the ‘best-bet’ neutron star merger detec- tion rates predicted for Advanced LIGO (Abadie et al. 2010), we estimate that up to about half a dozen of these transients should be detectable by ASKAP at any given time, although there are signiﬁcant uncer- tainties involved. Piran et al. (2012) independently es- timate ASKAP’s sensitivity to these events and obtain similar results. These are comparable to the rate es- timated from observations of short GRBs which have been proposed to originate in similar mergers (Coward et al. 2012). Metzger & Berger (2012) emphasise the chal- lenges related to uncertain GW event localisations, which are signiﬁcant. As discussed in §2.1.3, the large instantaneous sky coverage of ASKAP may be a sig- niﬁcant advantage in this situation.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Of the known neutron stars, the vast majority are pul- sars that have been detected through their pulsed radio emission. VAST will not be sensitive to radio pulsa- tions on timescales of . 1 second, but will instead probe neutron star populations that are not selected in pulsar surveys. We will be sensitive to intermittent or deeply nulling radio pulsars as well as radio pulsars with large variations in their average pulsed ﬂux.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Magnetars will be detectable through their rare ﬂaring events such as the giant ﬂare and expanding ra- dio nebula produced by SGR 1806−20 (Gaensler et al. 2005; Cameron et al. 2005). In addition, some magne- tars are known to turn on as radio sources. For exam- ple, XTE J1810−197 was initially detected as a tran- sient radio source (Halpern et al. 2005) before its ra- dio pulsations were identiﬁed (Camilo et al. 2006). We may also be able to detect bright intermittent pulses from the enigmatic rotating radio transients (RRATs; McLaughlin et al. 2006). VAST will thus reveal a broader census of the Galactic neutron star population than a standard pulsation search, with subsequent impli- cations for the overall supernova event rate in our Galaxy.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	The radio source FIRST J1023+0038 presents an interesting case study. Initially it was classiﬁed as a Galactic cataclysmic variable (CV) (Bond et al. 2002) because of its optical spectrum and variability, but its radio ﬂaring behavior appeared to be anomalous. The classiﬁcation was called into question because of the changes in the spectrum over time (e.g., Thorstensen & Armstrong 2005) and it was eventually re-identiﬁed as a transi- tional object where a recycled radio pulsar had turned on after a low mass X-ray binary phase (Archibald et al. 2009, 2010). The actual transition event had an ex- tremely short duration, with the accretion disk ap- parently dissipating between 2001 December and 2002 May. The radio pulsar was probably absent in 1998 (upper limits of 1.8 and 3.4 mJy at 1.4 GHz) but de- tected in a survey in mid-2007 with a ﬂux density ∼14 mJy at 1.6 GHz.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Novae occur in binaries (CVs) that consist of a white dwarf that is accreting matter from its companion star. A nova is a thermonuclear explosion in the layer of ac- creted matter on the surface of a white dwarf. A shell of matter is ejected during the explosion and this shell might produce radio emission. The radio emission can persist for a few years after the explosion, and in the case of FH Ser it reached ∼10 mJy at 2.7 GHz about one year after the outburst. In the late stages when the ejecta is optically thin, radio emission is generated by the entire ejecta and can therefore be used to es- timate the total ejecta mass (see e.g., Hjellming 1990 for a review).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	In AM Her-like systems (cataclysmic variables in which the white dwarf has a surface magnetic ﬁeld of ∼ 103 − 104 T) the M dwarfs are forced to co-rotate with the binary, which has an orbital period of a few hours. The dwarf star then becomes more active; ﬂares in AM Her itself have been observed to be as bright as 10 mJy at 5 GHz (Dulk et al. 1983).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Another group of stars that can produce radio ﬂares are the RS CVn-systems (Gunn et al. 1994). These are close detached binaries consisting of two main se- quence or evolved stars of spectral type F or later, that are tidally locked. The rapid spin of the stars leads to strong magnetic activity, which is manifested as large spots on the stars and also as ﬂaring activity that can be observed both in X-rays and at radio wavelengths. Flare-like outbursts are also observed in pre-main sequence stars, but it is not clear whether the physical mechanism is really that of a stellar ﬂare or related instead to the accretion ﬂow onto the star. Such out- bursts have been observed in both naked T Tauri stars, which do not show signs of extended accretion discs, and in systems with an infrared excess, which indicates the presence of a disc (Osten & Wolk 2009).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	A magnetic ﬁeld can convert the rotational energy of a star into an outﬂow and electromagnetic radiation. The most well-known examples of this are the classical radio pulsars, but a similar phenomenon can arise in less extreme stars. Radio observations have detected two fully circularly polarised radio pulses per spin pe- riod from the Ap star CU Vir, which is also spinning down in a way similar to a pulsar (e.g. Triglio et al. 2008; Trigilio et al. 2011; Lo et al. 2012).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Another example of activity driven by rotation is the white dwarf in the cataclysmic variable AE Aqr. The spin period of the white dwarf is 33 s, so rapid that the white dwarf does not accrete the matter that streams over from the companion K dwarf. Instead, the matter is ejected from the system. AE Aqr can be observed over a wide range of frequencies from X-rays to radio, and pulsed emission is observed at optical and X-rays, but it has not yet been detected at radio wave- lengths (De Jager et al. 1994; Bookbinder & Lamb 1987; Bastian et al. 1988).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Figure 2: The parameter space for radio tran- sients, adapted from Cordes et al. (2004). A quantity equivalent to absolute luminosity (ob- served ﬂux density S multiplied by the square of the distance D2) is plotted against the di- mensionless product of the emission frequency ν and the transient duration or pulse width W . In the Rayleigh-Jeans approximation, these quanti- ties are directly proportional and related to the brightness temperature, as indicated by the diago- nal lines, with T = 1012 K marking the maximum brightness temperature of incoherent processes. Examples of known radio transient sources are in- including giant pulses and “nanogiant” dicated, from the Crab pulsar pulses (Hankins et al. from XTE J1810−197 2003), radio pulses (Camilo et al. 2006) and other pulsars from the Australia Telescope National Facility pulsar cat- alogue (Manchester et al. 2005); the microquasar GRS 1915+105 (Mirabel & Rodr´ıguez 1994); ra- dio ﬂares from V4641 Sgr (Hjellming et al. 2000), the brown dwarf LP 944−20 (Berger et al. 2001), and the magnetar SGR 1806−20 (Gaensler et al. 2005; Cameron et al. 2005); the Galactic centre radio transient J1745−3009 (Hyman et al. 2005); pulses from the ultracool dwarf TVLM 513−46546 (Hallinan et al. 2007); as well as radio emission from the Sun and Jupiter. Red lines indicate the expected sensitivity of ASKAP to sources at dis- tances of 10 pc, 1 kpc and 1 Mpc.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	the last few decades. These surveys have been severely limited in their sensitivity, sky coverage, and cadence. In spite of this, they have revealed that the radio sky contains many transient objects, the identities of some of which remain mysterious.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	ing transient towards the Galactic Centre which lasted for only a few minutes with a ﬂux density in excess of 2 Jy at 330 MHz. Subsequent observations showed that the bursts repeated multiple times with a period of ∼ 77 minutes and a burst length of ∼ 10 minutes. The identiﬁcation of this source remains unclear (e.g. Kaplan et al. 2008; Roy et al. 2010), although similar- ities with bursts from ultracool dwarfs (Hallinan et al. 2007) are intruiging. Another example is the well- known Lorimer burst, discovered in an archival search of Parkes pulsar data. Lorimer et al. (2007) discovered a single millisecond pulse of extragalactic origin with the astonishing peak ﬂux density of 30 Jy at 1.4 GHz (also see Keane et al. 2011 for a second example). Al- though the nature of this source has been questioned, and recent work suggests the burst may be terrestrial (Burke-Spolaor et al. 2011), these new discoveries il- lustrate that the parameter space for radio variability has not been fully explored yet (Figure 2).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	In this section we summarise existing blind and archival surveys for radio transients, focusing on ‘slow’ transients with variability on timescales of seconds or greater.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Bower et al. (2007) presented the results from an archival survey of VLA data. Since the publication of this work there have been a number of blind radio transient sur- veys conducted on both archival data, and from the ﬁrst datasets produced by next generation radio tele- scopes. Most of these have had the dual aim of (a) searching for individual objects of interest and trying to identify them; and (b) characterising the population statistics of radio transients in order to prepare for fu- ture surveys. Although the reported transient rates from Bower et al. (2007) have recently been revised, after a reanalysis of the data by Frail et al. (2012), the work established a framework for comparison of radio transient surveys.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Equations 4 and 5 have been used to generate sur- vey predictions from the VAST survey parameters de- scribed in Section 4, and these predictions are shown in Figure 3. We also include the predicted rates of a sub- set of known radio transients (see Frail et al. 2012, for futher discussion). One limitation of the plot in Fig- ure 3 is that it collapses a number of dimensions into a single two-dimensional plot. For example, it does not represent the cadence of the surveys, or any trends in transient detection rate with observing frequency.
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Surveys typically fall into one of two categories: many repeated observations of a single ﬁeld (e.g. Carilli et al. 2003; Bower et al. 2007), or a large area, few-epoch surveys (e.g. Levinson et al. 2002; Gal-Yam et al. 2006). The two-epoch equivalent snapshot rate is a valid met- ric for quantising the eﬀectiveness of a survey, pro- vided the the timescale of transient behaviour is less than the cadence of the observations but longer than a single integration. In this case, a survey with many epochs is equivalent to a larger area survey with two epochs. Further survey metrics have been developed to account for observation cadence and source evolution timescales as well (e.g., Cordes 2007).
Processed_VAST:_An_ASKAP_Survey_for_Variables_and_Slow_Trans.txt	Table 3: Summary of snapshot rates for transient and variables radio sources reported in the literature. The results are organised according to upper limits based on non-detections (top section); transient detections (middle section); and detections of highly variable radio sources (bottom section).
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	Scientiﬁc Frontiers in Research on Extrasolar Planets ASP Conference Series, Vol. 28?, 2003 D. Deming et al.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	Abstract. We are setting up a new search for transiting extra-solar planets using the 0.5m Automated Patrol Telescope at Siding Spring Ob- servatory, Australia. We will begin regular observations in September 2002. We expect to ﬁnd ∼ 7 new planets per year.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	An increasing number of teams are searching for extra-solar planets using the transit method (see Horne 2002 for a review). Although the probability of observing a transit for any given star is small, using a wide-ﬁeld telescope a large number of stars can be monitored, potentially yielding a higher detection rate than the radial velocity surveys (e.g. Marcy et al. 2002; Mayor et al. 2002). Furthermore, for a transiting planet orbiting a suﬃciently bright star, the planet’s size, as well as its actual mass and orbital characteristics (from follow-up spectroscopy) can be determined, constraining models of its structure and formation.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	The Automated Patrol Telescope (APT) is a 0.5m telescope of Baker-Nunn design, modiﬁed for use with a CCD. It is owned and operated by the University of New South Wales, and located at Siding Spring Observatory, Australia. The current CCD camera images a 2◦×3◦ ﬁeld with 9.4′′ pixels. An upgrade is planned (for early 2003) to a pair of 3k×6k CCDs with 4.2′′ pixels, covering the entire useful ﬁeld of view of the telescope (∼ 6◦ in diameter). The telescope and dome are computer-controlled, with the possibility of remote or fully automated observation. About 50% of the observing time on the APT is reserved for this planet search.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	Figure 1. Photometric precision for a series of 150 second exposures taken with the APT. The ﬁeld is centered on the open cluster NGC6633, close to the Galactic plane. The limits due to photon shot noise in the star and sky ﬂux are shown. Readout noise is negligible. Systematic errors limit the precision for bright stars. Stars brighter than V ∼ 10.5 are saturated.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	intra-pixel sensitivity variations on the CCD, this causes photometric errors of ∼ 3% when using standard aperture photometry. Anderson and King (2000) have developed software which deals with the undersampling problem. We are using a modiﬁed version of this software. It builds up a model of the “eﬀective” point-spread function, which is the instrumental point-spread function convolved with the sensitivity of a single pixel, and measures stars by ﬁtting this model to the image. We can now measure stars down to V ∼ 14 with a relative precision of 1% (Fig 1) in a single 150 second exposure. For the brightest stars, systematic errors limit the precision. We are working on identifying and eliminating these errors.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	We are about to begin regular observations. We select four dense stellar ﬁelds near the Galactic plane and cycle between them, taking exposures through a V ﬁlter. This way we obtain 5 images of each ﬁeld per hour, for 6–8 hours per clear night. We observe each ﬁeld for 2–3 months.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	Kov´acs, Zucker, & Mazeh (2002) show that their box-ﬁtting algorithm for de- tecting transit signals in stellar lightcurves requires an eﬀective signal-to-noise ratio (for the combined measurements during the transit) of 6 for a signiﬁcant detection. For a typical 1% deep, 3 hour long transit, we can achieve this by observing just 3 transits. This is true for ∼ 1000 stars with V < 14 in one low Galactic latitude APT ﬁeld. To a magnitude limit of V = 14, near the Galactic plane, F, G, K and M type main sequence stars constitute ∼ 60% of the stars in the ﬁeld. Upper main sequence stars are too large for a Jupiter to cause a 1% transit. Thus, with the 5 times larger ﬁeld of view of our new CCD system, we will measure ∼ 3000 stars per ﬁeld with suﬃcient precision to detect an orbiting Jupiter.
Processed_An_Automated_Search_for_Extrasolar_Planet_Transits.txt	Due to observational constraints (weather, lunar and diurnal cycle, other projects), we estimate that we will detect 10% of transits which occur in our ﬁelds. We are most likely to detect close-orbiting planets with periods of a few days. Almost 1% of nearby stars host one of these “Hot Jupiters” (Butler et al. 2001). Their orbits have a ∼ 10% probability of being edge-on. We therefore expect to ﬁnd one planet per 104 stars measured with suﬃcient precision. We plan to observe 24 ﬁelds per year, yielding approximately 7 new planets each year.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	nuclear medium, one must also consider the structure of the mesons, as well as the nucleon.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	the scalar-meson mass is kept constant, and have shown the decrease of the hadron mass.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	of the nucleon, the mesons (σ, ω and ρ) and the hyperons (Λ, Σ and Ξ) in 40Ca and 208Pb.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	internal structure has time to adjust to the local ﬁelds, is naturally suited to the problem.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	constant scalar and vector ﬁelds, with values equal to those at the centre of the nucleon.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	CN would not depend on the scalar ﬁeld, that is, CN would be constant (CN = 1).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	so that the eﬀective σ-N coupling constant, gσ(σ), decreases at half the rate of CN (σ).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	with gσσ in MeV, and we introduce two parameters, aσ (in MeV−1) and bσ (in MeV−2).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	neutrons, and the total baryon density, ρB, is then given by ρp + ρn. Let the constant mean-ﬁeld values for the σ, ω and ρ ﬁelds be ¯σ, ¯ω and ¯b, respectively.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	mass be stationary under variation of the free bag radius (RN in the case of the nucleon).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	mass in matter (see Eq.(15)). (A) aσ = 3.0 × 10−4 (MeV−1) and bσ = 100 × 10−8 (MeV−2), (B) aσ = 5.0 × 10−4 (MeV−1) and bσ = 50 × 10−8 (MeV−2), (C) aσ = 7.5 × 10−4 (MeV−1) and bσ = 100 × 10−8 (MeV−2).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	that the nuclear compressibility is higher than that in QMC-I (K ∼ 200 – 300 MeV) [21].
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	the scalar ﬁeld, the eigenvalue is reduced more than 10% (at ρ0) from that in free space.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	parameters, we have calculated the masses of Λ, Σ and Ξ in symmetric nuclear matter.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	note that the dependence of aj on the hadrons is quite weak, and it ranges around 8.6 ∼ 9.5 × 10−4 (MeV−1).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	Table 3: Model parameters for ﬁnite nuclei (for mq = 5 MeV and RN = 0.8 fm).
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	In Fig.7 we ﬁrst show the calculated strength of the σ and ω ﬁelds in 40Ca and 208Pb.
Processed_Variation_of_hadron_masses_in_finite_nuclei.txt	the latter will certainly be important in any attempt to actually measure the mass shift.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Explaining and reasoning about processes which underlie observed black-box phenomena enables the discovery of causal mechanisms, derivation of suitable abstract representations and the for- mulation of more robust predictions. We pro- pose to learn high level functional programs in order to represent abstract models which cap- ture the invariant structure in the observed data. We introduce the π-machine (program-induction machine) – an architecture able to induce inter- pretable LISP-like programs from observed data traces. We propose an optimisation procedure for program learning based on backpropagation, gra- dient descent and A* search. We apply the pro- posed method to two problems: system identiﬁ- cation of dynamical systems and explaining the behaviour of a DQN agent. Our results show that the π-machine can efﬁciently induce inter- pretable programs from individual data traces.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	expect the system to do if a certain variable were changed or removed, or model checking (Baier & Katoen, 2008) of longer term properties including safety and large deviations in performance. We address these needs through a program induction based framework.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	We propose to learn high level functional programs in or- der to represent abstract models which capture the invariant structure in the observed data. Recent works have demon- strated the usefulness of program representations in cap- turing human-like concepts (Lake et al., 2015). Used in this way, program-based representations boost generalisa- tion and enable one-shot learning. Also, and arguably more importantly, they are signiﬁcantly more amenable to model checking and human interpretability.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	In this paper, we introduce the π-machine (program- induction machine), an architecture which is able to induce LISP-like programs from observed transition system data traces in order to explain various phenomena. Inspired by differentiable neural computers (Graves et al., 2014; 2016), the π-machine, as shown in Figure 1, is composed of a memory unit and a controller capable of learning programs from data by exploiting the scalability of stochastic gra- dient descent. However, the ﬁnal program obtained after training is not an opaque object encoded in the weights of a controller neural network, but a LISP-like program which provides a rigorous and interpretable description of the observed phenomenon. A key feature of our approach is that we allow the user to provide a set of predicates in order to specify the properties they are of interest interested in understanding as well as the context in which the data is to be explained. By exploiting the equivalence between computational graphs and functional programs we describe a hybrid optimisation procedure based on backpropagation, gradient descent, and A* search which is used to induce programs from data traces.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	We evaluate the performance of the π-machine on two dif- ferent problems. Firstly, we apply it to data from physics experiments and show that it is able to induce programs which represent fundamental laws of physics. The learn- ing procedure has access to relevant variables, but it does not have any other prior knowledge regarding physical laws which it has discovered in the same sense as in (Schmidt & Lipson, 2009) although far more computationally tractably.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Secondly, we study the use of the proposed procedure in explaining control policies learnt by a deep Q-network (DQN). Starting from behaviour traces of a reinforcement learning agent that has learnt to play the game of Pong, we demonstrate how the π-machine learns a functional pro- gram to describe that policy.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Explainability and interpretability. The immense suc- cess of deep neural network based learning systems and their rapid adoption in numerous real world application do- mains has renewed interest in the interpretability and ex- plainability of learnt models (Gunning, 2016). There is recognition that Bayesian rule lists (Letham et al., 2015; Yang et al., 2016), decision trees and probabilistic graphi- cal models are interpretable to the extent that they impose strong structural constraints on models of the observed data and allow for various types of queries, including introspec- tive and counterfactual ones. In contrast, deep learning models usually are trained ‘per query’ and have numerous parameters that could be hard to interpret. Zeiler & Fergus (2014) introduced deconvolutional networks in order to vi- sualise the layers of convolutional networks and provide a more intuitive understanding of why they perform well. Za- havy et al. (2016) describe Semi-Aggregated Markov Deci- sion Process (SAMDP) in order to analyse and understand the behaviour of a DQN based agent. Methods for tex- tual rationalisation of the predictions made by deep models have also been proposed (Harrison et al., 2017; Hendricks et al., 2016; Lei et al., 2016). While all of these works pro- vide useful direction, more generic methods are required which need not be hand-crafted to explain speciﬁc aspects of individual models. In this sense, we follow the model- agnostic explanation approach of Ribeiro et al. (2016), who provide “textual or visual artefacts” explaining the predic- tion of any classiﬁer by treating it as a black-box. Simi- larly to the way in which (Ribeiro et al., 2016) utilise local classiﬁers composed together to explain a more complex model, we present an approach to incrementally construct- ing functional programs that explain a complex transition system from more localised predicates of interest.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	as by Shapiro (1983) who used inductive inference in a logic programming setting, others have developed methods based on a variety of approaches ranging from SAT solvers (Solar-Lezama et al., 2006) to genetic algorithms (Schmidt & Lipson, 2009), which tend to scale poorly hence of- ten become restricted to a narrow class of programs. Re- cently, deep neural networks have been augmented with a memory unit resulting in models similar to the origi- nal von Neumann architecture. These models can induce programs through stochastic gradient descent by optimis- ing performance on input/output examples (Graves et al., 2014; 2016; Grefenstette et al., 2015) or synthetic execu- tion traces (Reed & De Freitas, 2015; Cai et al., 2017; Ling et al., 2017). Programs induced with such neural architec- tures are encoded in the parameters of the controller net- work and are, in general, not easily interpretable (particu- larly from the point of view of being able to ask counter- factual questions or performing model checking). Another approach is to directly generate the source code of the out- put program which yields consistent high level programs. Usually, these types of approaches require large amounts of labelled data - either program input/output examples (De- vlin et al., 2017; Balog et al., 2016) or input paired with the desired output program code (Yin & Neubig, 2017).
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Determining how many input/output examples or execution traces are required in order to generalise well is still an open research problem. However, in this paper, we focus atten- tion more on the explanatory power afforded by programs rather than on the broader problems of generalisation in the space of programs. While these characteristics are of course related, we take a view similar to that of (Ribeiro et al., 2016), arguing that it is possible to build from locally valid program fragments which provide useful insight into the black-box processes generating the data. By combin- ing gradient descent and A* search the π-machine is able to learn informative and interpretable high-level LISP-like programs, even just from a single observation trace.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	The π-machine treats the process which has generated the observed data as a black-box and attempts to induce LISP- like program which can be interpreted and used to explain the data. We show that the proposed method can be applied both to introspection of machine learning models and to the broader context of autonomous agents.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Consider the labelled transition system Ω(S, A, δ) where S is a non-empty set of states, A is a non-empty set of ac- tions, each parametrised by θ ∈ RD, and δ : S × A → S is the state transition function. We deﬁne an observa- tion trace T as a sequence of observed state-action pairs (st, at(θt)) ∈ S×A generated by the recursive relationship st+1 = δ (st, at( θt)) for 1 ≤ t ≤ T . We are interested in inducing a LISP-like functional program ρ which when ex- ecuted by an abstract machine is mapped to an execution trace Tρ such that Tρ and T are equivalent according to an input speciﬁcation.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Figure 1. Overall architecture of the π-machine. The current best candidate solution (1) is used to propose new, structurally more complex candidates (2). Each one of the new candidate programs ρ is optimised (3) through gradient descent by comparing its execution trace to the observation trace (4). The observation trace in this case is a demonstration of a tower building task. During execution, the program has access to memory (5) which stores both state variables and induced parameters. All new candidate programs are scored based on their performance and complexity and are inserted in the candidate solutions priority queue (6). Once the execution trace of a candidate matches the observation trace the ﬁnal solution is returned (7).
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	memory state conﬁgurations, I is the set of supported in- structions and ε : M × I → M speciﬁes the effect of each instruction. We consider two types of instructions – primi- tive actions which emulate the execution of a ∈ A or arith- metic functions f ∈ F such that I = A ∪ F. Furthermore, a set of observed state variables Mv ⊆ S, which vary over time, are stored in memory together with a set of induced free parameters Mp. The variables in Mv form a context which the program will be built on. A custom detector Dv, operating on the raw data stream, could be provided for each variable, thus enabling the user to make queries with respect to different contexts and property speciﬁcations.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	The proposed program induction procedure is based on two major steps. Firstly, we explain how a given functional pro- gram can be optimised such that the loss L(ρ) is minimised. Secondly, we explain how the space of possible program structures can be searched efﬁciently by utilising gradient information. An architectural overview of the π-machine is provided in Figure 1.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	The error function σact determines the difference between two actions, while σlen compares the lengths of the gen- erated and observed action traces. By providing the error functions σact and σlen one can target different aspects of the observation trace to be explained as they specify when two action traces are equivalent.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Neural networks are naturally expressed as computational graphs which are the most fundamental abstraction in com- putational deep learning frameworks (Tokui et al., 2015; Bergstra et al., 2010; Abadi et al., 2016). Optimisation within a computational graph is usually performed by push- ing the input through the entire graph in order to calculate the output (forward pass) and then backpropagating the er- ror signal to update each parameter (backward pass). A key observation for the development of the π-machine is that computational graphs and functional programs are equiva- lent as both describe arbitrary compositions of pure func- tions applied to input data. Therefore, similarly to a compu- tational graph, a functional program can also be optimised by executing the program (forward pass), measuring the er- ror signal and then performing backpropagation to update the program (backward pass).
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Forward executed instructions is it i1, . . . , in ∈ I which are executed by recursively calling ε(. . . ε(ε(M1, i1), i2) . . . , in). M1 is the initial memory state initialised with the observed variables from s1 and any induced parameters. The π-machine keeps a time counter t which is initialised to 1 and is automatically incremented whenever a primitive action instruction is executed. If the instruction ik is a primitive action, ik ∈ A, then the π-machine automatically sets ˆat = ik and invokes the error function σact(ˆat, ˆθt, at, θt), where ˆθt has been calculated by previous instructions. If the error is above a certain threshold emax the program execution is terminated and the backward pass is initiated. Otherwise, the time counter is incremented and the values of the variables in Mv are automatically updated to the new observed state. Essentially, the π-machine simulates the execution of each action reﬂecting any changes it has caused in the observed state. Alternatively, if the currently executed instruction ik is a function, ik ∈ F, then the resulting value is calculated and ik, together with its arguments, is added to a detailed call trace χ maintained by the π-machine. Importantly, each function argument is either a parameter or a variable read from memory at time t or the result of another function. All this information is kept in χ which eventually contains the computational tree of the program.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Gradient descent step. Once the gradient ∇pL(ρ) of the loss function with respect to each input parameter p ∈ xp is calculated we utilise AdaGrad (Duchi et al., 2011) to up- date the values of all parameters after each program exe- cution. The gradient ∇vL(ρ) with respect to each input variable v ∈ xv is also available. However, a variable cannot be simply updated in the direction of the gradient as it represents a symbol, not just a value. Variables can only take values from memory which is automatically up- dated according to the observation trace during execution. Nevertheless, the gradient provides important information about the direction of change which we utilise to ﬁnd vari- ables that minimise the loss. Whenever the memory state is automatically updated, a KD-tree is built for each type of variable stored in memory. We assume that the vari- ables in memory are real vectors with different length. So, we represent the KD-tree which stores all D-dimensional variables in memory at time t as KD t . If a d-dimensional variable v is to be optimised it is replaced with a temporary parameter ptemp initialised with vt which is the value of v read from memory at the respective time step t. The tempo- rary parameter ptemp is also updated with AdaGrad (Duchi et al., 2011). After each descent step, the nearest neighbour of the updated value p(cid:48) temp is determined by querying the KD-tree with Kd temp). If the result of the query is a dif- ferent d-dimensional variable u then the temporary param- eter is immediately set to ptemp = ut. As this often shifts the solution to a new region of the error space the gradi- ent history for all parameters p ∈ xp is reset. Eventually, when a solution is to be returned, the temporary parameters are substituted with their closest variables according to the respective KD t . The forward and backward passes are re- peated until the error is below the maximum error threshold emax or a maximum number of iterations is reached. After that the optimised program ρ∗ is scored according to its er- ror and complexity, and pushed to a priority queue holding potential solutions.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	We represent the space of possible program structures as a graph G = (T AST , E) where each node Ti ∈ T AST is a valid program abstract syntax tree (AST). There is an edge from Ti to Tj if and only if Tj can be obtained by replacing exactly one of the leaves in Ti with a subtree Ts of depth 1. The program induction procedure always starts with an empty program. So, we frame structure search as a path ﬁnding problem, solved through the use of A* search.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	by multiplying the corresponding Jacobian of f with the value of the variable at the time it was read from memory tr.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Figure 2. The π-machine explaining the behaviour of a pendulum (top) and a linear oscillator(bottom). The best 3 solutions for each system are shown in the middle.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	to the desired goal. The complexity function C(ρ) is the weighted sum of (i) maximum depth of the program AST; (ii) the number of free parameters; (iii) the number of variables used by the program; the weights of which we set to wC = [10, 5, 1]. These choices ensure that short programs, maximally exploiting structure of the observation trace, are preferred.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	The π-machine is implemented in Clojure, which is a LISP dialect supporting powerful data structures and homoiconic syntax. All experiments are run on an Intel Core i7-4790 processor with 32GB RAM and use the following list of functions, F: vector addition, subtraction and scaling.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Physical systems. Firstly, we apply the π-machine to model learning for physical systems. The transition dy- namics of a second order dynamical system is written as ¨x(t) = k1x(t) + k2 ˙x(t), where x(t) is the state of the system at time t and k1, k2 are system coefﬁcients. We have recreated an experiment described in Schmidt & Lip- son (2009), where the authors show the learning of physical laws associated with classical mechanical systems includ- ing the simple pendulum and linear oscillator. A diagram of these two systems is shown in Figure 2 (left). We set A = {accel(θ)} where θ ∈ R for both experiments. The observation trace for each system is generated by simulat- ing the dynamics for 1s at 100Hz. We specify the action error function as σact = (cid:107)ˆθ − θ(cid:107)2 and set σlen = 0. In both experiments x ∈ R and v = ˙x ∈ R represent linear position and velocity.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	After that, all possible replacement subtrees are con- structed. An AST subtree Ts of depth 1 represents a function call. We prune the number of possible functions in F by ensuring type consistency. Each leaf of Ts can be a parameter or a variable. So, all possible combinations are considered. New variable leaves are initialised to a random variable with suitable type from memory, while new parameter leaves are sampled from the multivariate normal distribution N (0, 0.1). As a result, if nf functions are type compatible with l and each function takes na arguments at most, then there are 2na · nf replacement subtrees, resulting in that many new candidates. All newly proposed candidates are optimised in parallel, scored by ftotal and pushed to the priority queue.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	The three best solutions found by the π-machine for each system are shown in Figure 2 (middle). The best solution for each system correctly represents the underlying laws of motion. The program describing the behaviour of the pen- dulum was induced in 18 iterations, while the linear oscil- lator program needed 146 iterations. The total number of possible programs with AST depth of 2, given the described experimental setup, is approximately 1.7 × 104. The aver- age duration of an entire iteration (propose new programs, optimise and evaluate) was 0.6s. Schmidt & Lipson (2009) achieve similar execution times, but distributed over 8 quad core computers (32 cores in total). The experimental re- sults demonstrate that the π-machine can efﬁciently induce programs representing fundamental laws of physics.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Figure 3. The π-machine explaining the behaviour of a DQN agent playing ATARI Pong. The best 3 solutions are shown in the middle.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Deep Q-network. This experiment is based on our view that the core deep neural network based policy learner and the explanation layer play complementary roles. There are numerous advantages to performing end-to-end pol- icy learning, such as DQN-learning from raw video, how- ever, there is also a need to explain the behaviour of the learnt policy with respect to user-deﬁned properties of in- terest. We consider explaining the behaviour of a DQN agent playing the ATARI Pong game and are interested in the question: how does the network control the position of the paddle in order to hit the ball when it is in the right side of the screen. A diagram of the experimental setup is shown in ﬁgure 3 (left). The behaviour of the DQN is observed during a single game. Since the environment is deterministic, the state transition function, which gener- ates the observation trace for this experiment, is the pol- icy π(s) that the DQN has learnt. We would like to ex- plain the behaviour of the DQN in terms of the position of the opponent, the ball and the DQN agent (so, not just in terms of RAM memory values, for instance). Therefore, the observation trace contains those positions which are ex- tracted from each frame by a predeﬁned detector. We set A = {move(θ)} where θ ∈ R and represent the discrete actions of the network left, right, nop as move(1), move(−1), move(0) respectively. We specify the action error function as σact = (cid:107)ˆθ − θ(cid:107)2 and set σlen = 0.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	dent that the DQN not only learns the value of each game state, but also the underlying dynamics of controlling the paddle. Furthermore, we have tested the performance of an agent following a greedy policy deﬁned by the induced program. In our experiments over 100 games this agent achieved a score of 11.1(±0.17). This is not quite the score of 18.9(±1.3) obtained by an optimised DQN, but it is bet- ter than human performance 9.3 (Mnih et al., 2015). This difference of course emanates from the predeﬁned detec- tor not capturing all aspects of what the perceptual layers in DQN have learnt, so improved detector choices should yield interpretable programs that also attain performance closer to the higher score of the black-box policy.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	The π-machine can be viewed as a framework for auto- matic network architecture design (Zoph & Le, 2017; Ne- grinho & Gordon, 2017), as different models can be ex- pressed as concise LISP-like programs. Deep learning methods for limiting the search space of possible programs, which poses the greatest challenge, have been proposed (Balog et al., 2016), but how they can be applied to more generic frameworks such as the π-machine is an open ques- tion. The speciﬁcation of variable detectors not only ad- dresses this issue, but enables the user to make targeted and well grounded queries about the observed data trace. Such detectors can also be learnt from raw data in an unsuper- vised fashion (Garnelo et al., 2016; Kim & Canny, 2017).
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	In conclusion, we propose a novel architecture, the π- machine, for inducing LISP-like functional programs from observed data traces by utilising backpropagation, stochas- tic gradient descent and A* search. The experimental re- sults demonstrate that the π-machine can efﬁciently induce interpretable programs from short data traces.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Abadi, Mart´ın, Agarwal, Ashish, Barham, Paul, Brevdo, Eugene, Chen, Zhifeng, Citro, Craig, Corrado, Greg S, Davis, Andy, Dean, Jeffrey, Devin, Matthieu, et al. Ten- sorﬂow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467, 2016.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Graves, Alex, Wayne, Greg, Reynolds, Malcolm, Harley, Tim, Danihelka, Ivo, Grabska-Barwi´nska, Agnieszka, Colmenarejo, Sergio G´omez, Grefenstette, Edward, Ra- malho, Tiago, Agapiou, John, et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):471–476, 2016.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Bergstra, James, Breuleux, Olivier, Bastien, Fr´ed´eric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guil- laume, Turian, Joseph, Warde-Farley, David, and Ben- gio, Yoshua. Theano: A cpu and gpu math compiler in python. In Proc. 9th Python in Science Conf, pp. 1–7, 2010.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Bottou, L´eon, Peters, Jonas, Candela, Joaquin Quinonero, Charles, Denis Xavier, Chickering, Max, Portugaly, Elon, Ray, Dipankar, Simard, Patrice Y, and Snelson, Ed. Counterfactual reasoning and learning systems: the ex- ample of computational advertising. Journal of Machine Learning Research, 14(1):3207–3260, 2013.
Processed_Using_Program_Induction_to_Interpret_Transition_Sy.txt	Grefenstette, Edward, Hermann, Karl Moritz, Suleyman, Mustafa, and Blunsom, Phil. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp. 1828–1836, 2015.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.