|<–Video for humans and machines||Towards a responsible AI–>|
The preceding chapters have dealt with the Moving Pictures and Audio part of the MPAI mission. However, MPAI does think that the benefits of AI can be extended to all types of data, not just those that can be directly consumed by machines. The purpose of this chapter is to present the state of the art in some of the data types being considered by MPAI, namely, financial data, online gaming, autonomous vehicles, and genomics.
One of the most important requests from the financial field is the ability to monitor the health of companies and detect evidence of anomalies to reduce the risk of future defaults and thus preserve business continuity. An extensive scientific literature about insolvency predictions was formulated in the late 1960s [12, 13] and during the first half of the 1980s [14, 15]. The purpose of the models reported in the literature is the identification of some indicators able to predict the level of risk and the possible default of the company by using appropriate econometric techniques.
Financial institutions, governments and in general the various market players have sought methods that are as efficient as possible and numerous researchers have pursued that goal by developing various quantitative methods, most of which are based on the statistical approach. Traditional models are accurate for about 12 months and in some cases, where the forecast has a sufficient accuracy, they are 70% precise in about 24 months [16, 17, 18, 19, 20]. This represents a severe limitation considering that the ideal forecasting model should allow medium-term predictions because the symptoms of a failure can be traced back to 5-8 years prior to failure. To overcome the characteristic limitations of statistical models, research work was carried out on pattern recognition methods developed in the field of ML. These studies have shown how ML models can offer better performance than traditional methods. Despite the greater accuracy, however, the ability to forecast in the medium term (over 24 months) remains, the main problem.
Therefore, research has focused on improving the forecasts’ accuracy and on extending the time horizon. Most of the literature has focused its efforts on selecting the most appropriate financial indicators and neglected non-financial information. On the other hand, the introduction of non-financial data could improve performance in terms of accuracy and the forecasting horizon, for both traditional and ML-based models .
Even though improvements have been recorded over the years, there is a gap between market demands and the best available practices. The current state of the art, in fact, does not offer models that are accurate for both short and medium term, versatile in relation to different markets with an agreed tuning method, possibly as an automated process and inclusive of the capability to analyse the effects on financial and non-financial variables.
In the history of online gaming, developers have always been confronted with the problems arising when information moves between the clients involved in the game and the server. They approached the problems with several strategies: by formally defining strategies and protocols that could solve the most obvious problems such as the unsynchronised and smooth display of actions between clients and server; by making protocols available to clients residing behind complex networks, designed for home connections and not in a local network with direct exposure of the player’s machine. After a series of evolutions and solutions such as Microsoft’s Direct Play and Gamespy, to date, they are still confronted with two problems that have not found a consolidated solution: the delay or absence of data packets, and cheating players.
There have long been attempts to solve the problem of missing data in a visual way from the client’s viewpoint by predicting the data that had not reached the system, based on the information available to the client at that moment. The prediction, however, was based on the data of the current game. So far, this situation sometimes still generates problems and inconsistencies.
Over time, different methods have been developed for different types of games in the case of cheating. These range from modifying the behaviour of the client by placing BOTs in charge of the more complex skills of the game to be managed (targeting opponents or making moves with very demanding timing) instead of the human player creating artificial delays in the delivery of packages to gain advantage during certain phases of the game. With the advent of authoritative online servers, however, the approach to cheating has changed. Since this type of architectural choice requires that the server always chooses and validates the actual game state, one possible way to cheat is to add visual aids to the client that give an advantage to the player not shared by other opponents. An example is provided by indicators that allow a player to immediately understand where the ball will end up in a football game.
Neural networks have already been used in several titles. The first experiments were made on classic games that could not find adequate AI models of computer opponents due to the very complexity of the game. One of these first scenarios was backgammon with TD-Gammon. For commercial video games, we find the use of neural networks in video games of different genres such as Electronic Arts’ Black & White (strategy game) and racing games.
In particular, the experience of Drivatar for the game “Forza Motorsport” made people realise how much data acquisition from the styles of community players could be used to build a computerized adversary that properly competes with the player. Starting in 2019, Milestone with its MotoGP franchise used the A.N.N.A. (Artificial Neural Network Agent) neural network. This technological solution has worked on the characters and skills of the riders so that they can be like their real-life counterparts and, at the same time, can adapt over time to the skill of the player.
An Autonomous Vehicle can move itself in the physical environment on the basis of high-level instructions received by humans or a machine and by processing data acquired from the environment. Connected Autonomous Vehicle (CAV) can send and receive data to/from other entities such as other CAVs and other devices, e.g., a traffic light of a roadside unit.
Some of the data types have an electromagnetic nature, namely:
- Global Navigation Satellite System (GNSS).
- Radio data from various sources and frequencies.
- Visual data in the human visible range (400-800 THz)
- Lidar data in the 200 THz range.
- Radar data in the 25 and 75 GHz range.
Other data types have an acoustic nature, namely:
- Ultrasound data in the 20 kHz range.
- Environmental audio in the audible range (16Hz-16 kHz).
Still other data have a heterogeneous nature, namely:
- Weather, air pressure, humidity, road conditions, etc.
- Position, Velocity and Acceleration.
The challenge for a CAV is the creation of an internal representation of the external world that is sufficiently robust to allow it to move itself to reach the instructed destination while satisfying a number of conditions that human drivers are assumed to know, for example, by passing appropriate examinations.
Considering that the automotive market is worth ~3.6 T$ in 20211 and the inevitable shift toward electric and eventually autonomous vehicles, it should not surprise us that many CAVs have been designed, built, and tested, and a few are being intensely trialled2.
The DNA of living beings is one of the most notable examples of natural data coding, evolved spontaneously to support life on Earth. Most cells of any living organism contain instructions that guide birth, growth, life, reproduction and interaction of an astronomical number of individuals belonging to a vast number of different species. The programs lying at the core of each living cell are expressed as long polymers of 4 different basic molecules (called “nucleotides”) sequentially attached at the side of very long strings of sugars. Interestingly, the cell’s programming is self-interpreting – it encodes the very same tools that will be used to decode it – and is specified in terms of a very abstract structure involving several different levels of regulation. So, at the most basic level, the cellular program specifies the production of nanomolecules and can read itself, extract energy from the environment and reproduce; however, the genomes of more complex life forms also encode information about how different cells interact among themselves to form complex organs and organisms, and implicit guidelines dictating how different organisms interact among themselves to form complex ecosystems. The programs can be remarkably complex – the DNA of each human cell, which is by no means the most complex of the genomes, has roughly 3.2 billion “letters” or nucleotides, equivalent to ~0.8 GBytes of binary encoded information. Considering that there are ~40 trillion cells in the human body, the total amount of data stored in each individual can be estimated at a staggering 32 EBytes (million TBytes). Each copy of the human genome stores blueprints for ~25,000 different types of nanomachines, plus a yet not well quantified or understood number of developmental programs that regulate the translation of the program into functional, well-adapted living creatures.
Humans may think that the DNA is very “human” and “personal”, but the data is not easily accessible for inspection. Costly equipment called sequencers can “read” DNA but doing so is technically challenging. The machines currently available output “noisy”, i.e., unreliable reads that are short fragments randomly extracted from much longer molecules without indication of their original placement. So, a sequencer must read the same genome many times, and provide many reads, to be reasonably sure of the value of a particular nucleotide – and reconstructing the original genome out of the short fragments is a challenging operation requiring huge computational resources. While human DNA has 3.2 billion nucleotides, the output of a sequencer must have a file size in the order of hundreds of GBytes to be reliably used. In addition, some parts of the genome are hard to sequence due to biological reasons – the Human Genome3 sequencing project is continuously updated since the production of the first draft human genome (or “reference”) in 2003.
Hundreds of algorithms and computer programs performing operations on the reads obtained from sequencers have been developed over the years. One vital problem is that of establishing how the genome of each human individual differs from the idealised human reference genome established by the Human Genome sequencing project, which does not correspond to any specific individual. While all human genomes share a very high level of sequence similarity, any two genomes differ by millions of small changes, which can be substitutions of a single DNA “letter” or more complex differences. Changes with respect to the human reference can confer desirable or deleterious traits – for instance, a single-nucleotide change in the blueprint of an essential protein can sometimes cause severe genetic disease. The ability to identify the specific characteristics of an individual’s DNA is making it possible to achieve what is called “personalised medicine” – genomic data can help discover whether an individual has an ongoing disease or risks developing one. By pinpointing the causes of an individual’s clinical status, genomic data processing can lead to more personalised treatments.
However, achieving such results is possible only if a suitable infrastructure is put in place. In addition to tools able to recognise the genomic variants characterising each individual and determine their clinical relevance, one typically needs strategies to encode, compress, store and access the genetic data output by sequencing machines. By using a zip-like compression function, a genomic file can be reduced by a factor of ~3-4. By applying smarter compression algorithms, the size of the file can be reduced by a much larger factor.
In general, making sense of genomic data is not easy, due to the number of hierarchically nested levels of regulation and the very abstract way information is encoded. As a result, progress in some critical areas can be very slow – and AI, with its ability to discover among vast amounts of data patterns otherwise hidden to humans, is being increasingly helpful. Recent progress with the problem of protein folding – i.e., the ability to computationally predict the 3D-structure of cellular components out of their linear blueprints stored in DNA – has made the headlines of newspapers and captured universal attention. Other typical applications are recommendation systems able to optimise therapies or treatment for individuals based on their genetic makeup.
1 Global Car & Automobile Sales – Market Size 2005–2027 https://www.ibisworld.com/global/market-size/global-car-automobile-sales/
2 Smart Mobility Projects and Trials Across the World, https://imoveaustralia.com/smart-mobility-projects-trials-list/
3 The Human Genome Project; https://www.genome.gov/human-genome-project
|<–Video for humans and machines||Towards a responsible AI–>|