PGF (Progressive Graphics File) is a wavelet-based bitmapped image format that employs lossless and lossy data compression. PGF was created to improve upon and replace the JPEG format. It was developed at the same time as JPEG 2000 but with a focus on speed over compression ratio. PGF can operate at higher compression ratios without taking more encoding/decoding time and without generating the characteristic "blocky and blurry" artifacts of the original DCT-based JPEG standard. It also allows more sophisticated progressive downloads. == Color models == PGF supports a wide variety of color models: Grayscale with 1, 8, 16, or 31 bits per pixel Indexed color with palette size of 256 RGB color image with 12, 16 (red: 5 bits, green: 6 bits, blue: 5 bits), 24, or 48 bits per pixel ARGB color image with 32 bits per pixel Lab color image with 24 or 48 bits per pixel CMYK color image with 32 or 64 bits per pixel == Technical discussion == PGF claims to achieve an improved compression quality over JPEG adding or improving features such as scalability. Its compression performance is similar to the original JPEG standard. Very low and very high compression rates (including lossless compression) are also supported in PGF. The ability of the design to handle a very large range of effective bit rates is one of the strengths of PGF. For example, to reduce the number of bits for a picture below a certain amount, the advisable thing to do with the first JPEG standard is to reduce the resolution of the input image before encoding it — something that is ordinarily not necessary for that purpose when using PGF because of its wavelet scalability properties. The PGF process chain contains the following four steps: Color space transform (in case of color images) Discrete Wavelet Transform Quantization (in case of lossy data compression) Hierarchical bit-plane run-length encoding === Color components transformation === Initially, images have to be transformed from the RGB color space to another color space, leading to three components that are handled separately. PGF uses a fully reversible modified YUV color transform. The transformation matrices are: [ Y r U r V r ] = [ 1 4 1 2 1 4 1 − 1 0 0 − 1 1 ] [ R G B ] ; [ R G B ] = [ 1 3 4 − 1 4 1 − 1 4 − 1 4 1 − 1 4 3 4 ] [ Y r U r V r ] {\displaystyle {\begin{bmatrix}Y_{r}\\U_{r}\\V_{r}\end{bmatrix}}={\begin{bmatrix}{\frac {1}{4}}&{\frac {1}{2}}&{\frac {1}{4}}\\1&-1&0\\0&-1&1\end{bmatrix}}{\begin{bmatrix}R\\G\\B\end{bmatrix}};\qquad \qquad {\begin{bmatrix}R\\G\\B\end{bmatrix}}={\begin{bmatrix}1&{\frac {3}{4}}&-{\frac {1}{4}}\\1&-{\frac {1}{4}}&-{\frac {1}{4}}\\1&-{\frac {1}{4}}&{\frac {3}{4}}\end{bmatrix}}{\begin{bmatrix}Y_{r}\\U_{r}\\V_{r}\end{bmatrix}}} The chrominance components can be, but do not necessarily have to be, down-scaled in resolution. === Wavelet transform === The color components are then wavelet transformed to an arbitrary depth. In contrast to JPEG 1992 which uses an 8x8 block-size discrete cosine transform, PGF uses one reversible wavelet transform: a rounded version of the biorthogonal CDF 5/3 wavelet transform. This wavelet filter bank is exactly the same as the reversible wavelet used in JPEG 2000. It uses only integer coefficients, so the output does not require rounding (quantization) and so it does not introduce any quantization noise. === Quantization === After the wavelet transform, the coefficients are scalar-quantized to reduce the amount of bits to represent them, at the expense of a loss of quality. The output is a set of integer numbers which have to be encoded bit-by-bit. The parameter that can be changed to set the final quality is the quantization step: the greater the step, the greater is the compression and the loss of quality. With a quantization step that equals 1, no quantization is performed (it is used in lossless compression). In contrast to JPEG 2000, PGF uses only powers of two, therefore the parameter value i represents a quantization step of 2i. Just using powers of two makes no need of integer multiplication and division operations. === Coding === The result of the previous process is a collection of sub-bands which represent several approximation scales. A sub-band is a set of coefficients — integer numbers which represent aspects of the image associated with a certain frequency range as well as a spatial area of the image. The quantized sub-bands are split further into blocks, rectangular regions in the wavelet domain. They are typically selected in a way that the coefficients within them across the sub-bands form approximately spatial blocks in the (reconstructed) image domain and collected in a fixed size macroblock. The encoder has to encode the bits of all quantized coefficients of a macroblock, starting with the most significant bits and progressing to less significant bits. In this encoding process, each bit-plane of the macroblock gets encoded in two so-called coding passes, first encoding bits of significant coefficients, then refinement bits of significant coefficients. Clearly, in lossless mode all bit-planes have to be encoded, and no bit-planes can be dropped. Only significant coefficients are compressed with an adaptive run-length/Rice (RLR) coder, because they contain long runs of zeros. The RLR coder with parameter k (logarithmic length of a run of zeros) is also known as the elementary Golomb code of order 2k. === Comparison with other file formats === JPEG 2000 is slightly more space-efficient in handling natural images. The PSNR for the same compression ratio is on average 3% better than the PSNR of PGF. It has a small advantage in compression ratio but longer encoding and decoding times. PNG (Portable Network Graphics) is more space-efficient in handling images with many pixels of the same color. There are several self-proclaimed advantages of PGF over the ordinary JPEG standard: Superior compression performance: The image quality (measured in PSNR) for the same compression ratio is on average 3% better than the PSNR of JPEG. At lower bit rates (e.g. less than 0.25 bits/pixel for gray-scale images), PGF has a much more significant advantage over certain modes of JPEG: artifacts are less visible and there is almost no blocking. The compression gains over JPEG are attributed to the use of DWT. Multiple resolution representation: PGF provides seamless compression of multiple image components, with each component carrying from 1 to 31 bits per component sample. With this feature there is no need for separately stored preview images (thumbnails). Progressive transmission by resolution accuracy, commonly referred to as progressive decoding: PGF provides efficient code-stream organizations which are progressive by resolution. This way, after a smaller part of the whole file has been received, it is possible to see a lower quality of the final picture, the quality can be improved monotonically getting more data from the source. Lossless and lossy compression: PGF provides both lossless and lossy compression in a single compression architecture. Both lossy and lossless compression are provided by the use of a reversible (integer) wavelet transform. Side channel spatial information: Transparency and alpha planes are fully supported ROI extraction: Since version 5, PGF supports extraction of regions of interest (ROI) without decoding the whole image. == Available software == The author published libPGF via a SourceForge, under the GNU Lesser General Public License version 2.0. Xeraina offers a free Windows console encoder and decoder, and PGF viewers based on WIC for 32bit and 64bit Windows platforms. Other WIC applications including File Explorer are able to display PGF images after installing this viewer. Digikam is a popular open-source image editing and cataloging software that uses libPGF for its thumbnails. It makes use of the progressive decoding feature of PGF images to store a single version of each thumbnail, which can then be decoded to different resolutions without loss, thus allowing users to dynamically change the size of the thumbnails without having to recalculate them again.
Google Tasks
Google Tasks is a task management application developed by Google and included with Google Workspace. Included initially as a feature in Gmail and Google Calendar, Google Tasks launched as a core product with a standalone app in 2018. It is available for Android and iOS, as well as in the right-hand side panel on Google Workspace apps on the web and in Google Calendar. == History and development == Google Tasks began as an integration within other apps in G Suite (now Google Workspace), allowing to-do items to be created in Calendar and Gmail. Upon graduating to a core service on June 28, 2018, Google Tasks launched as a dedicated mobile app in which tasks can be sorted into lists, managed, and completed. Google Tasks launched the ability to create tasks from Google Chat messages in 2022.
Pinakes
The Pinakes (Ancient Greek: Πίνακες 'tables', plural of πίναξ pinax) is a lost bibliographic work composed by Callimachus (310/305–240 BCE) that is popularly considered to be the first library catalog in the West; its contents were based upon the holdings of the Library of Alexandria during Callimachus's tenure there during the third century BCE. == History == The Library of Alexandria had been founded by Ptolemy I Soter about 306 BCE. The first recorded librarian was Zenodotus of Ephesus. During Zenodotus' tenure, Callimachus, who was never the head librarian, compiled many catalogues/lists, each called Pinakes. His most famous one listed authors and their works; thus he became the first known bibliographer and the scholar who organized the library by authors and subjects about 245 BCE. His work was 120 volumes long. Apollonius of Rhodes was the successor to Zenodotus. Eratosthenes of Cyrene succeeded Apollonius in 235 BCE and compiled his tetagmenos epi teis megaleis bibliothekeis, the 'scheme of the great bookshelves'. In 195 BCE Aristophanes of Byzantium, Eratosthenes' successor, was the librarian and updated the Pinakes, although it is also possible that his work was not a supplement of Callimachus' Pinakes themselves, but an independent polemic against, or commentary upon, their contents. == Description == The collection at the Library of Alexandria contained nearly 500,000 papyrus scrolls, which were grouped together by subject matter and stored in bins. Each bin carried a label with painted tablets hung above the stored papyri. Pinakes was named after these tablets and are a set of index lists. The bins gave bibliographical information for every roll. A typical entry started with a title and also provided the author's name, birthplace, father's name, any teachers trained under, and educational background. It contained a brief biography of the author and a list of the author's publications. The entry had the first line of the work, a summary of its contents, the name of the author, and information about the origin of the roll, as well as any doubts about the genuineness of the ascription. Callimachus' system divided works into six genres of poetry and five sections of prose: rhetoric, law, epic, tragedy, comedy, lyric poetry, history, medicine, mathematics, natural science, and miscellanies. Each category was alphabetized by author. Callimachus composed two other works that were referred as pinakes and were probably somewhat similar in format to the Pinakes (of which they "may or may not be subsections"), but were concerned with individual topics. These are listed by the Suda as: A Chronological Pinax and Description of Didaskaloi from the Beginning and Pinax of the Vocabulary and Treatises of Democritus. == Later bibliographic pinakes == The term pinax was used for bibliographic catalogs beyond Callimachus. For example, Ptolemy-el-Garib's catalog of Aristotle's writings comes to us with the title Pinax (catalog) of Aristotle's writings. == Legacy == The Pinakes proved indispensable to librarians for centuries, and they became a model for organizing knowledge throughout the Mediterranean. Their later influence can be traced to medieval times, even to the Arabic counterpart of the tenth century: Ibn al-Nadim's Al-Fihrist ("Index"). Local variations for cataloging and library classification continued through the late 19th century, when Anthony Panizzi and Melvil Dewey paved the way for more shared and standardized approaches.
Paradigms of AI Programming
Paradigms of AI Programming: Case Studies in Common Lisp (ISBN 1-55860-191-0) is a well-known programming book by Peter Norvig about artificial intelligence programming using Common Lisp. == History == The Lisp programming language has survived since 1958 as a primary language for artificial intelligence research. This text was published in 1992 as the Common Lisp standard was becoming widely adopted. Norvig introduces Lisp programming in the context of classic AI programs, including General Problem Solver (GPS) from 1959, ELIZA: Dialog with a Machine, from 1966, and STUDENT: Solving Algebra Word Problems, from 1964. The book covers more recent AI programming techniques, including Logic Programming, Object-Oriented Programming, Knowledge Representation, Symbolic Mathematics and Expert Systems.
TD-Gammon
TD-Gammon is a computer backgammon program developed in the 1990s by Gerald Tesauro at IBM's Thomas J. Watson Research Center. Its name comes from the fact that it is an artificial neural net trained by a form of temporal-difference learning, specifically TD-Lambda. It explored strategies that humans had not pursued and led to advances in the theory of correct backgammon play. In 1993, TD-Gammon (version 2.1) was trained with 1.5 million games of self-play, and achieved a level of play just slightly below that of the top human backgammon players of the time. In 1998, during a 100-game series, it was defeated by the world champion by a mere margin of 8 points. Its unconventional assessment of some opening strategies had been accepted and adopted by expert players. TD-gammon is commonly cited as an early success of reinforcement learning and neural networks, and was cited in, for example, papers for deep Q-learning and AlphaGo. == Algorithm for play and learning == During play, TD-Gammon examines on each turn all possible legal moves and all their possible responses (lookahead search), feeds each resulting board position into its evaluation function, and chooses the move that leads to the board position that got the highest score. In this respect, TD-Gammon is no different than almost any other computer board-game program. TD-Gammon's innovation was in how it learned its evaluation function. TD-Gammon's learning algorithm consists of updating the weights in its neural net after each turn to reduce the difference between its evaluation of previous turns' board positions and its evaluation of the present turn's board position—hence "temporal-difference learning". The score of any board position is a set of four numbers reflecting the program's estimate of the likelihood of each possible game result: White wins normally, Black wins normally, White wins a gammon, Black wins a gammon. For the final board position of the game, the algorithm compares with the actual result of the game rather than its own evaluation of the board position. The core of TD-gammon is a neural network with 3 layers. The input layer has two types of neurons. One type codes for the board position. They are non-negative integers ranging from 0 to 15, indicating the number of White or Black checkers at each board location. There are 99 input neurons for each, totaling 198 neurons. Another type codes for hand-crafted features previously used in Neurogammon. These features encoded standard concepts used by human experts, such as "advanced anchor," "blockade strength," "home board strength" and the probability of a "blot" (single checker) being hit. The hidden layer contains hidden neurons. Later versions had more of these. The output layer contains 4 neurons, representing the network's estimate of the probability ("equity") that the current board would lead to. The 4 neurons code for: White normal win, White gammon win, Black normal win, Black gammon win. Backgammon win is so rare that Tesauro opted to not represent it. After each turn, the learning algorithm updates each weight in the neural net according to the following rule: w t + 1 − w t = α ( Y t + 1 − Y t ) ∑ k = 1 t λ t − k ∇ w Y k {\displaystyle w_{t+1}-w_{t}=\alpha (Y_{t+1}-Y_{t})\sum _{k=1}^{t}\lambda ^{t-k}\nabla _{w}Y_{k}} where: It was found that picking small λ {\displaystyle \lambda } offered performance roughly equally good, and large λ {\displaystyle \lambda } degraded performance. Because of this, after 1992, TD-Gammon was trained with λ = 0 {\displaystyle \lambda =0} , degenerating into standard TD-learning. This saved compute by a factor of 2. == Development history == Version 1.0 used simple 1-ply search: every next move is scored by the neural net, and the highest-scoring move is selected. Versions 2.0 and 2.1 used 2-ply search: Make a 1-ply analysis to remove unlikely moves ("forward pruning"). Make a 2-play minimax analysis for only the likely moves. Pick the best move, probability-weighted by each of the opponent's 21 possible dice rolls (weighting non-doubles twice as much as doubles). Versions 3.0 and 3.1 used 3-ply search, using 21 2 = 441 {\displaystyle 21^{2}=441} possible dice rolls instead of 21. The last version, 3.1, was trained specifically for an exhibition match against Malcolm Davis at the 1998 AAAI Hall of Champions. It lost at -8 points, mainly due to one blunder, where TD-Gammon opted to double and got gammoned at -32 points. == Experiments and stages of training == Unlike previous neural-net backgammon programs such as Neurogammon (also written by Tesauro), where an expert trained the program by supplying the "correct" evaluation of each position, TD-Gammon was at first programmed "knowledge-free". In early experimentation, using only a raw board encoding with no human-designed features, TD-Gammon reached a level of play comparable to Neurogammon: that of an intermediate-level human backgammon player. Even though TD-Gammon discovered insightful features on its own, Tesauro wondered if its play could be improved by using hand-designed features like Neurogammon's. Indeed, the self-training TD-Gammon with expert-designed features soon surpassed all previous computer backgammon programs. It stopped improving after about 1,500,000 games (self-play) using a three-layered neural network, with 198 input units encoding expert-designed features, 80 hidden units, and one output unit representing predicted probability of winning. == Advances in backgammon theory == TD-Gammon's exclusive training through self-play (rather than imitation learning) enabled it to explore strategies that humans previously had not considered or had ruled out erroneously. Its success with unorthodox strategies had a significant impact on the backgammon community. Late 1991, Bill Robertie, Paul Magriel, and Malcolm Davis, were invited to play against TD-Gammon (version 1.0). A total of 51 games were played, with TD-Gammon losing at -0.25 ppg. Robertie found TD-Gammon to be at the level of a competent advanced player, and better than any previous backgammon program. Robertie subsequently wrote about the use of TD-Gammon for backgammon study. For example, on the opening play, the conventional wisdom was that given a roll of 2-1, 4-1, or 5-1, White should move a single checker from point 6 to point 5. Known as "slotting", this technique trades the risk of a hit for the opportunity to develop an aggressive position. TD-Gammon found that the more conservative play of splitting 24-23 was superior. Tournament players began experimenting with TD-Gammon's move, and found success. Within a few years, slotting had disappeared from tournament play, replaced by splitting, though in 2006 it made a reappearance for 2-1. Backgammon expert Kit Woolsey found that TD-Gammon's positional judgement, especially its weighing of risk against safety, was superior to his own or any human's. TD-Gammon's excellent positional play was undercut by occasional poor endgame play. The endgame requires a more analytical approach, sometimes with extensive lookahead. TD-Gammon's limitation to two-ply lookahead put a ceiling on what it could achieve in this part of the game. TD-Gammon's strengths and weaknesses were the opposite of symbolic artificial intelligence programs and most computer software in general: it was good at matters that require an intuitive "feel" but bad at systematic analysis. It is also poor at doubling strategies. This is likely due to the fact that the neural network is trained without the doubling cube, with the doubling added by feeding the neural network's cubeless equity estimates into theoretically-based heuristic formulae. This was particularly the case in the 1998 exhibition match, where it played 100 games against Malcolm Davis. A single doubling blunder lost the match. TD-gammon was never commercialized or released to the public in some other form, but it inspired commercial backgammon programs based on neural networks, such as JellyFish (1994) and Snowie (1998).
E-on Vue
Vue is a software tool for world generation by Bentley Systems, with support for many visual effects, animations, and various other features. The tool has been used in several feature-length films. In 2024, Bentley Systems announced that Vue would be discontinued, and be freely available to those that still wish to use it. == Versions == == Features == This is a list of features as of the 2023 release of Vue: === Terrains === Heightfield terrains Procedural terrains Infinite terrains Planetary terrains Real-world terrains 3D terrain sculpting Terrain export === EcoSystem Instancing Technology === Material-based EcoSystems Global EcoSystems Dynamic EcoSystems 360° EcoSystem Population Paint EcoSystem instances EcoParticles Export EcoSystem populations === Vegetation === Built-in Plant editor Compatible with PlantFactory Vegetation assets === Atmosphere, Skies and Clouds === Standard atmospheric model Spectral atmospheric model Photometric atmospheric model Atmosphere presets Procedural Volumetric 3D cloud layers Standalone 3D Metaclouds Convert meshes to Clouds Cloud morphing Import OpenVDB Export standalone and cloud layer zones to OpenVDB Export skies as HDRI === Modeling === Primitive and Feature modeling 3D Text edition tool Metablobbing Hyperblobs Export baked hyperblobs Splines Built in Road Construction toolkit Random rock generator Export rocks === Texturing and UVs === Material presets PBR Substance support Node-based procedural materials Volumetric materials and Hypertextures Stacked UVs Unwrapped UVs Ptex === Interoperability, Integration And Export === Export single assets to generic 3D formats Full scene export Integration plugins Import and Export Camera data as FBX and Nuke.chan Python API ZBrush GoZ bridge === Animation === Animate objects, materials, atmospheres, clouds, waves... Automatic wind and breeze Localized wind effects per plant / per EcoSystem population Omni and directional ventilators for local modifications of plants Time spline editor Automatic keyframe creation Automatic synchronization of cameras and lights Animation export as AfterEffects Import motion tracking information === Lighting === Global illumination, Global Radiosity, Ambient occlusion Subsurface Scattering HDRI image based lighting Point light, Quadratic point light, Spotlight, Quadratic spotlight, Directional light Use IES distribution profiles on photometric lights Area lights, light panels, light portals Physically accurate caustics computation === Rendering === Render with Ray Tracer Render with Path Tracer Stereoscopic rendering 360/180 VR Panorama Render Option Spherical panoramic rendering Tone mapping options Multipass & G-Buffer Network rendering with HyperVue / RenderCows Network rendering with RenderNodes == Users == Blue Sky Studios Digital Domain DreamWorks Animation: Kung Fu Panda Industrial Light & Magic: Indiana Jones and the Kingdom of the Crystal Skull, Pirates of the Caribbean: Dead Man's Chest Sony Pictures Imageworks Warner Bros. Interactive Entertainment Weta Digital
TD-Gammon
TD-Gammon is a computer backgammon program developed in the 1990s by Gerald Tesauro at IBM's Thomas J. Watson Research Center. Its name comes from the fact that it is an artificial neural net trained by a form of temporal-difference learning, specifically TD-Lambda. It explored strategies that humans had not pursued and led to advances in the theory of correct backgammon play. In 1993, TD-Gammon (version 2.1) was trained with 1.5 million games of self-play, and achieved a level of play just slightly below that of the top human backgammon players of the time. In 1998, during a 100-game series, it was defeated by the world champion by a mere margin of 8 points. Its unconventional assessment of some opening strategies had been accepted and adopted by expert players. TD-gammon is commonly cited as an early success of reinforcement learning and neural networks, and was cited in, for example, papers for deep Q-learning and AlphaGo. == Algorithm for play and learning == During play, TD-Gammon examines on each turn all possible legal moves and all their possible responses (lookahead search), feeds each resulting board position into its evaluation function, and chooses the move that leads to the board position that got the highest score. In this respect, TD-Gammon is no different than almost any other computer board-game program. TD-Gammon's innovation was in how it learned its evaluation function. TD-Gammon's learning algorithm consists of updating the weights in its neural net after each turn to reduce the difference between its evaluation of previous turns' board positions and its evaluation of the present turn's board position—hence "temporal-difference learning". The score of any board position is a set of four numbers reflecting the program's estimate of the likelihood of each possible game result: White wins normally, Black wins normally, White wins a gammon, Black wins a gammon. For the final board position of the game, the algorithm compares with the actual result of the game rather than its own evaluation of the board position. The core of TD-gammon is a neural network with 3 layers. The input layer has two types of neurons. One type codes for the board position. They are non-negative integers ranging from 0 to 15, indicating the number of White or Black checkers at each board location. There are 99 input neurons for each, totaling 198 neurons. Another type codes for hand-crafted features previously used in Neurogammon. These features encoded standard concepts used by human experts, such as "advanced anchor," "blockade strength," "home board strength" and the probability of a "blot" (single checker) being hit. The hidden layer contains hidden neurons. Later versions had more of these. The output layer contains 4 neurons, representing the network's estimate of the probability ("equity") that the current board would lead to. The 4 neurons code for: White normal win, White gammon win, Black normal win, Black gammon win. Backgammon win is so rare that Tesauro opted to not represent it. After each turn, the learning algorithm updates each weight in the neural net according to the following rule: w t + 1 − w t = α ( Y t + 1 − Y t ) ∑ k = 1 t λ t − k ∇ w Y k {\displaystyle w_{t+1}-w_{t}=\alpha (Y_{t+1}-Y_{t})\sum _{k=1}^{t}\lambda ^{t-k}\nabla _{w}Y_{k}} where: It was found that picking small λ {\displaystyle \lambda } offered performance roughly equally good, and large λ {\displaystyle \lambda } degraded performance. Because of this, after 1992, TD-Gammon was trained with λ = 0 {\displaystyle \lambda =0} , degenerating into standard TD-learning. This saved compute by a factor of 2. == Development history == Version 1.0 used simple 1-ply search: every next move is scored by the neural net, and the highest-scoring move is selected. Versions 2.0 and 2.1 used 2-ply search: Make a 1-ply analysis to remove unlikely moves ("forward pruning"). Make a 2-play minimax analysis for only the likely moves. Pick the best move, probability-weighted by each of the opponent's 21 possible dice rolls (weighting non-doubles twice as much as doubles). Versions 3.0 and 3.1 used 3-ply search, using 21 2 = 441 {\displaystyle 21^{2}=441} possible dice rolls instead of 21. The last version, 3.1, was trained specifically for an exhibition match against Malcolm Davis at the 1998 AAAI Hall of Champions. It lost at -8 points, mainly due to one blunder, where TD-Gammon opted to double and got gammoned at -32 points. == Experiments and stages of training == Unlike previous neural-net backgammon programs such as Neurogammon (also written by Tesauro), where an expert trained the program by supplying the "correct" evaluation of each position, TD-Gammon was at first programmed "knowledge-free". In early experimentation, using only a raw board encoding with no human-designed features, TD-Gammon reached a level of play comparable to Neurogammon: that of an intermediate-level human backgammon player. Even though TD-Gammon discovered insightful features on its own, Tesauro wondered if its play could be improved by using hand-designed features like Neurogammon's. Indeed, the self-training TD-Gammon with expert-designed features soon surpassed all previous computer backgammon programs. It stopped improving after about 1,500,000 games (self-play) using a three-layered neural network, with 198 input units encoding expert-designed features, 80 hidden units, and one output unit representing predicted probability of winning. == Advances in backgammon theory == TD-Gammon's exclusive training through self-play (rather than imitation learning) enabled it to explore strategies that humans previously had not considered or had ruled out erroneously. Its success with unorthodox strategies had a significant impact on the backgammon community. Late 1991, Bill Robertie, Paul Magriel, and Malcolm Davis, were invited to play against TD-Gammon (version 1.0). A total of 51 games were played, with TD-Gammon losing at -0.25 ppg. Robertie found TD-Gammon to be at the level of a competent advanced player, and better than any previous backgammon program. Robertie subsequently wrote about the use of TD-Gammon for backgammon study. For example, on the opening play, the conventional wisdom was that given a roll of 2-1, 4-1, or 5-1, White should move a single checker from point 6 to point 5. Known as "slotting", this technique trades the risk of a hit for the opportunity to develop an aggressive position. TD-Gammon found that the more conservative play of splitting 24-23 was superior. Tournament players began experimenting with TD-Gammon's move, and found success. Within a few years, slotting had disappeared from tournament play, replaced by splitting, though in 2006 it made a reappearance for 2-1. Backgammon expert Kit Woolsey found that TD-Gammon's positional judgement, especially its weighing of risk against safety, was superior to his own or any human's. TD-Gammon's excellent positional play was undercut by occasional poor endgame play. The endgame requires a more analytical approach, sometimes with extensive lookahead. TD-Gammon's limitation to two-ply lookahead put a ceiling on what it could achieve in this part of the game. TD-Gammon's strengths and weaknesses were the opposite of symbolic artificial intelligence programs and most computer software in general: it was good at matters that require an intuitive "feel" but bad at systematic analysis. It is also poor at doubling strategies. This is likely due to the fact that the neural network is trained without the doubling cube, with the doubling added by feeding the neural network's cubeless equity estimates into theoretically-based heuristic formulae. This was particularly the case in the 1998 exhibition match, where it played 100 games against Malcolm Davis. A single doubling blunder lost the match. TD-gammon was never commercialized or released to the public in some other form, but it inspired commercial backgammon programs based on neural networks, such as JellyFish (1994) and Snowie (1998).