Kórsafn (Icelandic: Choral archives) is a sound installation by Icelandic artist Björk. Developed in collaboration with the technology company Microsoft, audio design firm Listen and architecture office firm Atelier Ace, the installation was designed for the lobby of the Sister City Hotel in New York City, United States, and launched in 2020. Elaborating 17 years of choral recording taken from Björk discography, Kórsafn consisted of an evolving music composition that uses an artificial intelligence model that responds to real-time weather data, creating a continuously shifting auditory experience. == Background and concept == In 2018, Björk announced her tenth concert tour Cornucopia, which debuted as a residency show at The Shed arts center. Before the start of the show, it was confirmed she would be accompanied by The Hamrahlid Choir. In 2019, while she was performing at The Shed, Björk stayed alongside the choir at the Sister City Hotel in New York City, where they would rehearse for the performances. While there, the Atelier Ace, which owns the Sister City boutique hotels, asked her to create a sound installation for the lobby. This was the second work commissioned by the hotel, a year after a similar piece by Julianna Barwick was featured in the lobby. Kórsafn is formed from two Icelandic words, "kór" ("choral") and "safn" ("archives"). The installation features recordings of Björk’s choral works from the previous 17 years, including compositions taken from her albums Medúlla (2004) and Biophilia (2011). The artificial intelligence system was developed in collaboration with Microsoft. The software processes data gathered from sensors and by a camera placed on the roof of the Sister City Hotel building and by a barometer. It then uses algorithms to determine how the choral elements are layered, pitched, and mixed in real time. The AI generate variations in real time by reacting to the passage of flocks, clouds, airplanes and changes in pressure. Data collected from sensors on the hotel’s rooftop include wind speed, cloud cover, and precipitation levels. These inputs influence the tonal quality, volume, and rhythmic patterns of the soundscape. The sound is played through hidden speakers in the hotel's lobby, blending with the architectural environment to create an immersive experience for guests. The AI system learns over time from the changing of the seasons and weather constantly evolving the sound - keeping in harmony with the sky. Björk described the project as an "AI tango," expressing curiosity about the interplay between her choral compositions and the AI's interpretations of environmental data. She noted the significance of the Hudson Valley's rich bird migrations, which influence the generative aspects of the soundscape. Due to the COVID-19 pandemic, the hotel closed while the installation was ongoing, making a version of the sound piece available online. == Reception == Kórsafn was positively reviewed. It's Nice That author Jenny Brewer described the piece as "a high-tech alternative to the smooth jazz that usually whistles through hotel lobbies". Writing for CNET, Scott Stein observed that it "is lovely and low-key, and honestly, it just blends into the background. It's nothing wild, but it fits the hotel", adding that "after an hour, it didn't get annoying, or too repetitive". The installation garnered several recognitions. It was nominated in the Fast Company's 2020 Innovation by Design Awards in the Hospitality category. It received three Clio Awards silver prizes, in the Use of Music in Experience/Activation, Sound Design and Emerging Technology categories.
Kuwahara filter
The Kuwahara filter is a non-linear smoothing filter used in image processing for adaptive noise reduction. Most filters that are used for image smoothing are linear low-pass filters that effectively reduce noise but also blur out the edges. However the Kuwahara filter is able to apply smoothing on the image while preserving the edges. It is named after Michiyoshi Kuwahara, Ph.D., who worked at Kyoto and Osaka Sangyo Universities in Japan, developing early medical imaging of dynamic heart muscle in the 1970s and 80s. == The Kuwahara operator == Suppose that I ( x , y ) {\displaystyle I(x,y)} is a grey scale image and that we take a square window of size 2 a + 1 {\displaystyle 2a+1} centered around a point ( x , y ) {\displaystyle (x,y)} in the image. This square can be divided into four smaller square regions Q i = 1 ⋯ 4 {\displaystyle Q_{i=1\cdots 4}} each of which will be Q i ( x , y ) = { [ x , x + a ] × [ y , y + a ] if i = 1 [ x − a , x ] × [ y , y + a ] if i = 2 [ x − a , x ] × [ y − a , y ] if i = 3 [ x , x + a ] × [ y − a , y ] if i = 4 {\displaystyle Q_{i}(x,y)={\begin{cases}\left[x,x+a\right]\times \left[y,y+a\right]&{\mbox{ if }}i=1\\\left[x-a,x\right]\times \left[y,y+a\right]&{\mbox{ if }}i=2\\\left[x-a,x\right]\times \left[y-a,y\right]&{\mbox{ if }}i=3\\\left[x,x+a\right]\times \left[y-a,y\right]&{\mbox{ if }}i=4\\\end{cases}}} where × {\displaystyle \times } is the cartesian product. Pixels located on the borders between two regions belong to both regions so there is a slight overlap between subregions. The arithmetic mean m i ( x , y ) {\displaystyle m_{i}(x,y)} and standard deviation σ i ( x , y ) {\displaystyle \sigma _{i}(x,y)} of the four regions centered around a pixel (x,y) are calculated and used to determine the value of the central pixel. The output of the Kuwahara filter Φ ( x , y ) {\displaystyle \Phi (x,y)} for any point ( x , y ) {\displaystyle (x,y)} is then given by Φ ( x , y ) = m i ( x , y ) {\textstyle \Phi (x,y)=m_{i}(x,y)} where i = a r g min j σ j ( x , y ) {\displaystyle i=\operatorname {arg\min } _{j}\sigma _{j}(x,y)} . This means that the central pixel will take the mean value of the area that is most homogenous. The location of the pixel in relation to an edge plays a great role in determining which region will have the greater standard deviation. If for example the pixel is located on a dark side of an edge it will most probably take the mean value of the dark region. On the other hand, should the pixel be on the lighter side of an edge it will most probably take a light value. On the event that the pixel is located on the edge it will take the value of the more smooth, least textured region. The fact that the filter takes into account the homogeneity of the regions ensures that it will preserve the edges while using the mean creates the blurring effect. Similarly to the median filter, the Kuwahara filter uses a sliding window approach to access every pixel in the image. The size of the window is chosen in advance and may vary depending on the desired level of blur in the final image. Bigger windows typically result in the creation of more abstract images whereas small windows produce images that retain their detail. Typically windows are chosen to be square with sides that have an odd number of pixels for symmetry. However, there are variations of the Kuwahara filter that use rectangular windows. Additionally, the subregions do not need to overlap or have the same size as long as they cover all of the window. == Color images == For color images, the filter should not be performed by applying the filter to each RGB channel separately, and then recombining the three filtered color channels to form the filtered RGB image. The main problem with that is that the quadrants will have different standard deviations for each of the channels. For example, the upper left quadrant may have the lowest standard deviation in the red channel, but the lower right quadrant may have the lowest standard deviation in the green channel. This situation would result in the color of the central pixel to be determined by different regions, which might result in color artifacts or blurrier edges. To overcome this problem, for color images a slightly modified Kuwahara filter must be used. The image is first converted into another color space, the HSV color space. The modified filter then operates on only the "brightness" channel, the Value coordinate in the HSV model. The variance of the "brightness" of each quadrant is calculated to determine the quadrant from which the final filtered color should be taken from. The filter will produce an output for each channel which will correspond to the mean of that channel from the quadrant that had the lowest standard deviation in "brightness". This ensures that only one region will determine the RGB values of the central pixel. ImageMagick uses a similar approach, but using the Rec. 709 Luma as the brightness metric. === Julia Implementation === == Applications == Originally the Kuwahara filter was proposed for use in processing RI-angiocardiographic images of the cardiovascular system. The fact that any edges are preserved when smoothing makes it especially useful for feature extraction and segmentation and explains why it is used in medical imaging. The Kuwahara filter however also finds many applications in artistic imaging and fine-art photography due to its ability to remove textures and sharpen the edges of photographs. The level of abstraction helps create a desirable painting-like effect in artistic photographs especially in the case of the colored image version of the filter. These applications have known great success and have encouraged similar research in the field of image processing for the arts. Although the vast majority of applications have been in the field of image processing there have been cases that use modifications of the Kuwahara filter for machine learning tasks such as clustering. The Kuwahara filter has been implemented in CVIPtools. The Kuwahara filter is present as a shader node in Blender. == Drawbacks and restrictions == The Kuwahara filter despite its capabilities in edge preservation has certain drawbacks. At a first glance it is noticeable that the Kuwahara filter does not take into account the case where two regions have equal standard deviations. This is not often the case in real images since it is rather hard to find two regions with exactly the same standard deviation due to the noise that is always present. In cases where two regions have similar standard deviations the value of the center pixel could be decided at random by the noise in these regions. Again this would not be a problem if the regions had the same mean. However, it is not unusual for regions of very different means to have the same standard deviation. This makes the Kuwahara filter susceptible to noise. Different ways have been proposed for dealing with this issue, one of which is to set the value of the center pixel to ( m 1 + m 2 ) / 2 {\textstyle (m_{1}+m_{2})/2} in cases where the standard deviation of two regions do not differ more than a certain value D {\displaystyle D} . The Kuwahara filter is also known to create block artifacts in the images especially in regions of the image that are highly textured. These blocks disrupt the smoothness of the image and are considered to have a negative effect in the aesthetics of the image. This phenomenon occurs due to the division of the window into square regions. A way to overcome this effect is to take windows that are not rectangular(i.e. circular windows) and separate them into more non-rectangular regions. There have also been approaches where the filter adapts its window depending on the input image. == Extensions of the Kuwahara filter == The success of the Kuwahara filter has spurred an increase the development of edge-enhancing smoothing filters. Several variations have been proposed for similar use most of which attempt to deal with the drawbacks of the original Kuwahara filter. The "Generalized Kuwahara filter" proposed by P. Bakker considers several windows that contain a fixed pixel. Each window is then assigned an estimate and a confidence value. The value of the fixed pixel then takes the value of the estimate of the window with the highest confidence. This filter is not characterized by the same ambiguity in the presence of noise and manages to eliminate the block artifacts. The "Mean of Least Variance"(MLV) filter, proposed by M.A. Schulze also produces edge-enhancing smoothing results in images. Similarly to the Kuwahara filter it assumes a window of size 2 d − 1 × 2 d − 1 {\displaystyle 2d-1\times 2d-1} but instead of searching amongst four subregions of size d × d {\displaystyle d\times d} for the one with minimum variance it searches amongst all possible d × d {\displaystyle d\times d} subregions. This means the central pixel of the window will be assigned the mean of the one subregion out of a poss
Region connection calculus
The region connection calculus (RCC) is intended to serve for qualitative spatial representation and reasoning. RCC abstractly describes regions (in Euclidean space, or in a topological space) by their possible relations to each other. RCC8 consists of 8 basic relations that are possible between two regions: disconnected (DC) externally connected (EC) equal (EQ) partially overlapping (PO) tangential proper part (TPP) tangential proper part inverse (TPPi) non-tangential proper part (NTPP) non-tangential proper part inverse (NTPPi) From these basic relations, combinations can be built. For example, proper part (PP) is the union of TPP and NTPP. == Axioms == RCC is governed by two axioms. for any region x, x connects with itself for any region x, y, if x connects with y, y connects with x == Remark on the axioms == The two axioms describe two features of the connection relation, but not the characteristic feature of the connect relation. For example, we can say that an object is less than 10 meters away from itself and that if object A is less than 10 meters away from object B, object B will be less than 10 meters away from object A. So, the relation 'less-than-10-meters' also satisfies the above two axioms, but does not talk about the connection relation in the intended sense of RCC. == Composition table == The composition table of RCC8 are as follows: "" denotes the universal relation, no relation can be discarded. Usage example: if a TPP b and b EC c, (row 4, column 2) of the table says that a DC c or a EC c. == Examples == The RCC8 calculus is intended for reasoning about spatial configurations. Consider the following example: two houses are connected via a road. Each house is located on an own property. The first house possibly touches the boundary of the property; the second one surely does not. What can we infer about the relation of the second property to the road? The spatial configuration can be formalized in RCC8 as the following constraint network: house1 DC house2 house1 {TPP, NTPP} property1 house1 {DC, EC} property2 house1 EC road house2 { DC, EC } property1 house2 NTPP property2 house2 EC road property1 { DC, EC } property2 road { DC, EC, TPP, TPPi, PO, EQ, NTPP, NTPPi } property1 road { DC, EC, TPP, TPPi, PO, EQ, NTPP, NTPPi } property2 Using the RCC8 composition table and the path-consistency algorithm, we can refine the network in the following way: road { PO, EC } property1 road { PO, TPP } property2 That is, the road either overlaps (PO) property2, or is a tangential proper part of it. But, if the road is a tangential proper part of property2, then the road can only be externally connected (EC) to property1. That is, road PO property1 is not possible when road TPP property2. This fact is not obvious, but can be deduced once we examine the consistent "singleton-labelings" of the constraint network. The following paragraph briefly describes singleton-labelings. First, we note that the path-consistency algorithm will also reduce the possible properties between house2 and property1 from { DC, EC } to just DC. So, the path-consistency algorithm leaves multiple possible constraints on 5 of the edges in the constraint network. Since each of the multiple constraints involves 2 constraints, we can reduce the network to 32 (25) possible unique constraint networks, each containing only single labels on each edge ("singleton labelings"). However, of the 32 possible singleton labelings, only 9 are consistent. (See qualreas for details.) Only one of the consistent singleton labelings has the edge road TPP property2 and the same labeling includes road EC property1. Other versions of the region connection calculus include RCC5 (with only five basic relations - the distinction whether two regions touch each other are ignored) and RCC23 (which allows reasoning about convexity). == RCC8 use in GeoSPARQL == RCC8 has been partially implemented in GeoSPARQL as described below: == Implementations == GQR is a reasoner for RCC-5, RCC-8, and RCC-23 (as well as other calculi for spatial and temporal reasoning) qualreas is a Python framework for qualitative reasoning over networks of relation algebras, such as RCC-8, Allen's interval algebra and more.
OpenVINO
OpenVINO is an open-source software toolkit developed by Intel for optimizing and deploying deep learning models. It supports several popular model formats and categories, such as large language models, computer vision, and generative AI. OpenVINO is optimized for Intel hardware, but offers support for ARM/ARM64 processors. It sees great use in AI Sound Processing drivers when tied with Intel's Gaussian & Neural Accelerator (GNA). Based in C++, it extends API support for C and Python, as well as Node.js (in early preview). OpenVINO is cross-platform and free for use under Apache License 2.0. == Workflow == The simplest OpenVINO usage involves obtaining a model and running it as is. Yet for the best results, a more complete workflow is suggested: obtain a model in one of supported frameworks, convert the model to OpenVINO IR using the OpenVINO Converter tool, optimize the model, using training-time or post-training options provided by OpenVINO's NNCF. execute inference, using OpenVINO Runtime by specifying one of several inference modes. == OpenVINO model format == OpenVINO IR is the default format used to run inference. It is saved as a set of two files, .bin and .xml, containing weights and topology, respectively. It is obtained by converting a model from one of the supported frameworks, using the application's API or a dedicated converter. Models of the supported formats may also be used for inference directly, without prior conversion to OpenVINO IR. Such an approach is more convenient but offers fewer optimization options and lower performance, since the conversion is performed automatically before inference. Some pre-converted models can be found in the Hugging Face repository. The supported model formats are: PyTorch TensorFlow TensorFlow Lite ONNX (including formats that may be serialized to ONNX) PaddlePaddle JAX/Flax == OS support == OpenVINO runs on Windows, Linux and MacOS.
Darkforest
Darkforest is a computer go program developed by Meta Platforms, based on deep learning techniques using a convolutional neural network. Its updated version Darkfores2 combines the techniques of its predecessor with Monte Carlo tree search. The MCTS effectively takes tree search methods commonly seen in computer chess programs and randomizes them. With the update, the system is known as Darkfmcts3. Darkforest is of similar strength to programs like CrazyStone and Zen. It has been tested against a professional human player at the 2016 UEC cup. Google's AlphaGo program won against a professional player in October 2015 using a similar combination of techniques. Darkforest is named after Liu Cixin's science fiction novel The Dark Forest. == Background == Competing with top human players in the ancient game of Go has been a long-term goal of artificial intelligence. Go's high branching factor makes traditional search techniques ineffective, even on cutting-edge hardware, and Go's evaluation function could change drastically with one stone change. However, by using a Deep Convolutional Neural Network designed for long-term predictions, Darkforest has been able to substantially improve the win rate for bots over more traditional Monte Carlo Tree Search based approaches. === Matches === Against human players, Darkfores2 achieves a stable 3d ranking on KGS Go Server, which roughly corresponds to an advanced amateur human player. However, after adding Monte Carlo Tree Search to Darkfores2 to create a much stronger player named darkfmcts3, it can achieve a 5d ranking on the KGS Go Server. ==== Against other AI ==== darkfmcts3 is on par with state-of-the-art Go AIs such as Zen, DolBaram and Crazy Stone, but lags behind AlphaGo. It won 3rd place in January 2016 KGS Bot Tournament against other Go AIs. === News coverage === After Google's AlphaGo won against Fan Hui in 2015, Facebook made its AI's hardware designs public, alongside releasing the code behind DarkForest as open-source, in addition to heavy recruiting to strengthen its team of AI engineers. == Style of play == Darkforest uses a neural network to sort through the 10100 board positions, and find the most powerful next move. However, neural networks alone cannot match the level of good amateur players or the best search-based Go engines, and so Darkfores2 combines the neural network approach with a search-based machine. A database of 250,000 real Go games were used in the development of Darkforest, with 220,000 used as a training set and the rest used to test the neural network's ability to predict the next moves played in the real games. This allows Darkforest to accurately evaluate the global state of the board, but local tactics were still poor. Search-based engines have poor global evaluation, but are good at local tactics. Combining these two approaches is difficult because search-based engines work much faster than neural networks, a problem which was solved in Darkfores2 by running the processes in parallel with frequent communication between the two. === Conventional strategies === Go is generally played by analyzing the position of the stones on the board. Various advanced players have described it as playing in some part subconsciously. Unlike chess and checkers, where AI players can simply look further forward at moves than human players, but with each round of Go having on average 250 possible moves, that approach is ineffective. Instead, neural networks copy human play by training the AI systems on images of successful moves, the AI can effectively learn how to interpret how the board looks, as many grandmasters do. In November 2015, Facebook demonstrated the combination of MCTS with neural networks, which played with a style that "felt human". === Flaws === It has been noted that Darkforest still has flaws in its playstyle. The bot sometimes plays tenuki ("move elsewhere") pointlessly when local powerful moves are required. When the bot is losing, it shows the typical behavior of MCTS, it plays bad moves and loses more. The Facebook AI team has acknowledged these as areas of future improvement. == Program architecture == The family of Darkforest computer go programs is based on convolution neural networks. The most recent advances in Darkfmcts3 combined convolutional neural networks with more traditional Monte Carlo tree search. Darkfmcts3 is the most advanced version of Darkforest, which combines Facebook's most advanced convolutional neural network architecture from Darkfores2 with a Monte Carlo tree search. Darkfmcts3 relies on a convolution neural networks that predicts the next k moves based on the current state of play. It treats the board as a 19x19 image with multiple channels. Each channel represents a different aspect of board information based upon the specific style of play. For standard and extended play, there are 21 and 25 different channels, respectively. In standard play, each players liberties are represented as six binary channels or planes. The respective plane is true if the player one, two, or three or more liberties available. Ko (i.e. illegal moves) is represented as one binary plane. Stone placement for each opponent and empty board positions are represented as three binary planes, and the duration since a stone has been placed is represented as real numbers on two planes, one for each player. Lastly, the opponents rank is represented by nine binary planes, where if all are true, the player is a 9d level, if 8 are true, an 8d level, and so forth. Extended play additionally considers the border (binary plane that is true at the border), position mask (represented as distance from the board center, i.e. x ( − 0.5 ∗ d i s t a n c e 2 ) {\displaystyle x^{(-0.5distance^{2})}} , where x {\displaystyle x} is a real number at a position), and each player's territory (binary, based on which player a location is closer to). Darkfmct3 uses a 12-layer full convolutional network with a width of 384 nodes without weight sharing or pooling. Each convolutional layer is followed by a rectified linear unit, a popular activation function for deep neural networks. A key innovation of Darkfmct3 compared to previous approaches is that it uses only one softmax function to predict the next move, which enables the approach to reduce the overall number of parameters. Darkfmct3 was trained against 300 random selected games from an empirical dataset representing different game stages. The learning rate was determined by vanilla stochastic gradient descent. Darkfmct3 synchronously couples a convolutional neural network with a Monte Carlo tree search. Since the convolutional neural network is computationally taxing, the Monte Carlo tree search focuses computation on the more likely game play trajectories. By running the neural network synchronously with the Monte Carlo tree search, it is possible to guarantee that each node is expanded by the moves predicted by the neural network. == Comparison with other systems == Darkfores2 beats Darkforest, its neural network-only predecessor, around 90% of the time, and Pachi, one of the best search-based engines, around 95% of the time. On the Kyu rating system, Darkforest holds a 1-2d level. Darkfores2 achieves a stable 3d level on KGS Go Server as a ranked bot. With the added Monte Carlo tree search, Darkfmcts3 with 5,000 rollouts beats Pachi with 10k rollouts in all 250 games; with 75k rollouts it achieves a stable 5d level in KGS server, on par with state-of-the-art Go AIs (e.g., Zen, DolBaram, CrazyStone); with 110k rollouts, it won the 3rd place in January KGS Go Tournament.
Label noise
Label noise refers to errors or inaccuracies in the class labels of data instances. This is a widespread issue in machine learning datasets, arising from human annotator mistakes, unclear labeling instructions, automated labeling methods, or adversarial attacks in supervised learning. Label noise can be roughly divided into random noise, where labels are flipped independently of input features, and systematic noise, where mislabeling is dependent on certain patterns or biases in the data. Label noise can be damaging to model performance, especially for complex models that may overfit to noisy labels rather than generalizable patterns. Many approaches have been proposed to deal with the effects of label noise, including robust loss functions, noise-tolerant algorithms, data cleaning methods, and semi-supervised learning approaches. To reduce the impact of wrong labels during training, techniques like label smoothing, sample reweighting and using trusted validation sets are used. The role of noise-robust training paradigms and curriculum learning strategies to improve resilience against mislabeled data is also explored in recent research.
OpenAI Five
OpenAI Five is a computer program by OpenAI that plays the five-on-five video game Dota 2. Its first public appearance occurred in 2017, where it was demonstrated in a live one-on-one game against the professional player Dendi, who lost to it. The following year, the system had advanced to the point of performing as a full team of five, and began playing against and showing the capability to defeat professional teams. By choosing a game as complex as Dota 2 to study machine learning, OpenAI thought they could more accurately capture the unpredictability and continuity seen in the real world, thus constructing more general problem-solving systems. The algorithms and code used by OpenAI Five were eventually borrowed by another neural network in development by the company, one which controlled a physical robotic hand. OpenAI Five has been compared to other similar cases of artificial intelligence (AI) playing against and defeating humans, such as AlphaStar in the video game StarCraft II, AlphaGo in the board game Go, Deep Blue in chess, and Watson on the television game show Jeopardy!. == History == Development on the algorithms used for the bots began in November 2016. OpenAI decided to use Dota 2, a competitive five-on-five video game, as a base due to it being popular on the live streaming platform Twitch, having native support for Linux, and had an application programming interface (API) available. Before becoming a team of five, the first public demonstration occurred at The International 2017 in August, the annual premiere championship tournament for the game, where Dendi, a Ukrainian professional player, lost against an OpenAI bot in a live one-on-one matchup. After the match, CTO Greg Brockman explained that the bot had learned by playing against itself for two weeks of real time, and that the learning software was a step in the direction of creating software that can handle complex tasks "like being a surgeon". OpenAI used a methodology called reinforcement learning, as the bots learn over time by playing against itself hundreds of times a day for months, in which they are rewarded for actions such as killing an enemy and destroying towers. By June 2018, the ability of the bots expanded to play together as a full team of five and were able to defeat teams of amateur and semi-professional players. At The International 2018, OpenAI Five played in two games against professional teams, one against the Brazilian-based paiN Gaming and the other against an all-star team of former Chinese players. Although the bots lost both matches, OpenAI still considered it a successful venture, stating that playing against some of the best players in Dota 2 allowed them to analyze and adjust their algorithms for future games. The bots' final public demonstration occurred in April 2019, where they won a best-of-three series against The International 2018 champions OG at a live event in San Francisco. A four-day online event to play against the bots, open to the public, occurred the same month. There, the bots played in 42,729 public games, winning 99.4% of those games. == Architecture == Each OpenAI Five bot is a neural network containing a single layer with a 4096-unit LSTM that observes the current game state extracted from the Dota developer's API. The neural network conducts actions via numerous possible action heads (no human data involved), and every head has meaning. For instance, the number of ticks to delay an action, what action to select – the X or Y coordinate of this action in a grid around the unit. In addition, action heads are computed independently. The AI system observes the world as a list of 20,000 numbers and takes an action by conducting a list of eight enumeration values. Also, it selects different actions and targets to understand how to encode every action and observe the world. OpenAI Five has been developed as a general-purpose reinforcement learning training system on the "Rapid" infrastructure. Rapid consists of two layers: it spins up thousands of machines and helps them 'talk' to each other and a second layer runs software. By 2018, OpenAI Five had played around 180 years worth of games in reinforcement learning running on 256 GPUs and 128,000 CPU cores, using Proximal Policy Optimization, a policy gradient method. == Comparisons with other game AI systems == Prior to OpenAI Five, other AI versus human experiments and systems have been successfully used before, such as Jeopardy! with Watson, chess with Deep Blue, and Go with AlphaGo. In comparison with other games that have used AI systems to play against human players, Dota 2 differs as explained below: Long run view: The bots run at 30 frames per second for an average match time of 45 minutes, which results in 80,000 ticks per game. OpenAI Five observes every fourth frame, generating 20,000 moves. By comparison, chess usually ends before 40 moves, while Go ends before 150 moves. Partially observed state of the game: Players and their allies can only see the map directly around them. The rest of it is covered in a fog of war which hides enemies units and their movements. Thus, playing Dota 2 requires making inferences based on this incomplete data, as well as predicting what their opponent could be doing at the same time. By comparison, Chess and Go are "full-information games", as they do not hide elements from the opposing player. Continuous action space: Each playable character in a Dota 2 game, known as a hero, can take dozens of actions that target either another unit or a position. The OpenAI Five developers allow the space into 170,000 possible actions per hero. Without counting the perpetual aspects of the game, there are an average of ~1,000 valid actions each tick. By comparison, the average number of actions in chess is 35 and 250 in Go. Continuous observation space: Dota 2 is played on a large map with ten heroes, five on each team, along with dozens of buildings and non-player character (NPC) units. The OpenAI system observes the state of a game through developers' bot API, as 20,000 numbers that constitute all information a human is allowed to get access to. A chess board is represented as about 70 lists, whereas a Go board has about 400 enumerations. == Reception == OpenAI Five have received acknowledgement from the AI, tech, and video game community at large. Microsoft founder Bill Gates called it a "big deal", as their victories "required teamwork and collaboration". Chess champion Garry Kasparov, who lost against the Deep Blue AI in 1997, stated that despite their losing performance at The International 2018, the bots would eventually "get there, and sooner than expected". In a conversation with MIT Technology Review, AI experts also considered OpenAI Five system as a significant achievement, as they noted that Dota 2 was an "extremely complicated game", so even beating non-professional players was impressive. PC Gamer wrote that their wins against professional players was a significant event in machine learning. In contrast, Motherboard wrote that the victory was "basically cheating" due to the simplified hero pools on both sides, as well as the fact that bots were given direct access to the API, as opposed to using computer vision to interpret pixels on the screen. The Verge wrote that the bots were evidence that the company's approach to reinforcement learning and its general philosophy about AI was "yielding milestones". In 2019, DeepMind unveiled a similar bot for StarCraft II, AlphaStar. Like OpenAI Five, AlphaStar used reinforcement learning and self-play. The Verge reported that "the goal with this type of AI research is not just to crush humans in various games just to prove it can be done. Instead, it's to prove that — with enough time, effort, and resources — sophisticated AI software can best humans at virtually any competitive cognitive challenge, be it a board game or a modern video game." They added that the DeepMind and OpenAI victories were also a testament to the power of certain uses of reinforcement learning. It was OpenAI's hope that the technology could have applications outside of the digital realm. In 2018, they were able to reuse the same reinforcement learning algorithms and training code from OpenAI Five for Dactyl, a human-like robot hand with a neural network built to manipulate physical objects. In 2019, Dactyl solved the Rubik's Cube.