Personalized, stylized, specialized and localized machine intelligent training method

Abstract

The invention discloses a personalized, stylized, specialized and localized machine intelligence training method, and belongs to the field of machine intelligence and artificial intelligence. According to the method, only experience created by a single individual of an intelligent behavior body is taken as a training material, complete end-to-end training is carried out on machine intelligence, so that the neural network machine intelligence simulates the individual of the intelligent behavior body, and the trained machine intelligence shows technical skills, experience consciousness, style preference and rhythm pace of the individual of the intelligent behavior body; the method has the advantages that the method has natural internal consistency, reconstructs' my intelligent branch ', is especially interested in cites, organic division of labor and tacit cooperation in cooperation with the method, is more targeted for specific professional fields or specific space-time areas, and consumers can select a certain specific intelligent behavior individual to independently train machine intelligence; reputations or brands which are tested for a long time or are widely welcomed will be selected and inherited, and iterative extrapolation training can deal with new situations with times.

Description

Technical Field

[0001] The present invention relates to the fields of machine intelligence and artificial intelligence. Background Art

[0002] The so-called "intelligent actor" refers to the subject that behaves due to intelligence. It can be a human individual, an organization formed by a human collective, a machine intelligent body where machine intelligence resides, or a human in a machine, or a machine in a human, or a human-machine community where humans and machines co-drive, and others. The essence of many behavioral activities of intelligent actors is creation, not mechanical engineering, but a behavioral process that is different each time and moves towards and ultimately achieves the goal under the guidance of purpose, constraints of rules, and interaction with the environment. Machine intelligence can imitate this behavioral process. For example, for human driving vehicles or intelligent driving of smart cars, even if the driving tasks are exactly the same, the specific process of completing the same driving task each time is different, the action rhythm may be different, the driving route may be different, and the impact of other traffic participants may also be different. Therefore, such a driving experience process is actually a creation under the guidance of task purpose, constraints of traffic rules, and interaction with changing environments. Such experience data can be used as training materials for training machine intelligence.

[0003] The so-called "end-to-end" intelligence refers to the intelligent form of task result output directly from the output end by inputting raw data from the input end without going through the intermediate stages, details, logic, and concepts. For example, a baby reads "book" when seeing an object similar to a book. Such intelligent form can come from "end-to-end" training, that is, the overall training process of one intelligent actor training another intelligent actor. This kind of training builds and optimizes the model based on the raw data. For example, parents pick up a book and tell the baby "this is read as book" while repeatedly reading the sound of "book" themselves, and the baby's The input is "eyes see parents gesture "book" and "ears hear parents pronounce "book"", and the output is "one's own mouth pronounces "book". The repetition of this behavioral training enables the baby to establish and optimize the complex system model of its internal long chain giant network. From then on, when the baby sees a book-like object, he or she will have the ability to pronounce "book" in the tone, intonation, rhythm and accent of the parents. This training does not teach the specific concept of book, nor does it teach the key points of the oral muscles. It does not require the baby to understand the grammar of the sentence "this is pronounced book" said by the parents, nor does it require this sentence to have consistency in the same language. Instead, it focuses on overall "end-to-end" behavioral imitation.

[0004] In the training of machine intelligence, the industry often adopts "end-to-end" training based on deep learning of neural networks. For machine intelligence such as smart driving and smart painting, which mainly reach and ultimately achieve their goals through behavioral processes, the so-called "end-to-end" training means that the training materials of this machine intelligence during training are the original input and output experience data such as the purpose, state, environment, and behavior of the learning object. The neural network model is established and optimized through training. After the training is completed, in the actual application of the model, namely the machine operation of smart driving and smart painting, the original data such as the purpose, state, environment, etc. of the actual application are input, and the imitative behavioral data are directly output to the output end. Taking intelligent driving as an example, the neural network machine intelligence trained "end-to-end" can use the real driving experience data created by humans (such as the collection of driving purpose data, vehicle status data, driving situation video data and its corresponding acceleration, sliding, braking, steering and other maneuvering action horizontal and vertical driving control data) as training materials. Through training, the intelligent driving neural network model is established and optimized. After the training is completed, the model can achieve the output from the vehicle driving status and environmental video input end directly to the vehicle maneuvering action horizontal and vertical driving control data output end under the guidance of clear task objectives, that is, "clear task objectives, state environment input, driving control behavior output". The calculation program does not go through the traditional staged logical links such as perception, prediction, planning, decision-making, and control, nor does it have specific qualitative variable programming for concepts such as "traffic lights", "lane lines", and "pedestrians". That is, such machine intelligence and training methods focus on holistic imitation of the behavior process of the learning object (goal orientation, rule constraints, and movement and ultimate achievement of the goal under environmental interaction). Only the input and output data are identifiable, and the neural network in the middle is a black box.

[0005] In the current machine intelligence and artificial intelligence industry, for complete "end-to-end" training, almost all manufacturers and practitioners use behavioral representation data created by multiple intelligent actors (learning objects). This is because, in general, the training materials created by a single intelligent actor are relatively small, which is difficult to meet the needs of complete "end-to-end" training for massive training materials. At the same time, the speed at which a single intelligent actor creates training materials is slow, which is difficult to meet the needs of manufacturers and practitioners to complete training in the shortest possible time. More importantly, the industry generally believes or expects that the use of experience data created by multiple intelligent actors (learning objects) can eliminate the individuality of experience data. The long-term goal of moving towards universality is to make machine intelligence move towards the so-called "general artificial intelligence". This is a technical bias that has gradually formed in the industry. It blindly and unfoundedly expects that machine intelligence will one day have general intelligence like humans, which can draw inferences from one instance and apply it to other instances. It makes the entire industry not think about the benefits and welfare that personalized, stylized, specialized and localized machine intelligence may bring to humans. It deviates from the objective fact that humans need personalized, stylized, specialized and localized machine intelligence in some fields, and especially ignores the general demand that human users strongly hope to have another "my intelligent avatar" (rather than "general artificial intelligence") to help them with their work in some machine intelligence usage scenarios. Under such technical bias, the use of experience data created by multiple intelligent actors (learning objects) in the machine intelligence industry for complete "end-to-end" training is almost the only common phenomenon. For example, a certain end-to-end neural network intelligent driving manufacturer has 12 test drivers around the world, located in New Zealand, Thailand, Norway and Japan, collecting driving data of its test vehicles driving around the world as training materials, instead of using the driving data of one tester or one test vehicle as training materials, in order to meet its FSD Beta The end-to-end training of the V12 autonomous driving model requires a large amount of data. During the training process, 10 million screened and labeled human driving videos are input. Its fleet of nearly 2 million vehicles around the world provides about 160 billion frames of original video materials for training every day. An end-to-end neural network visual artificial intelligence manufacturer employs a labeling team of dozens of people to manually mark the visual objects in its training materials, rather than just one person. An end-to-end neural network content generation chatbot manufacturer uses the corpus of many content contributors in knowledge websites as training materials, rather than just the corpus of a certain person. An end-to-end neural network content generation poetry robot manufacturer uses the poems of multiple poets in a certain dynasty as training materials, rather than just the poems of a certain poet.

[0006] The disadvantages are that multiple learning objects are multiple different intelligent actors, each with its own style preferences and rhythm pace (the unique correspondence between "the order, amplitude, weight, change of behavior" and "time"). In the process of completing the same task, some are radical and some are conservative, some are impatient and some are slow, some focus on speed and some focus on quality, some have a fast operating rhythm and some have a slow operating rhythm, some are slow at first and then fast, and some are fast at first and then slow. Their different style preferences lead to large differences in the distribution of experience data, resulting in poor fitting effect of the training result model; the machine intelligence trained by the mixed experience data of multiple different intelligent actors is mediocre and styleless, and the information representing the style is mixed and mediocrely disappears in the experience data of multiple different intelligent actors, and does not express the individual style of a specific intelligent actor. The individual experience of some excellent intelligent actors is difficult to be inherited, and it is difficult to meet the personalized and stylized needs of consumers. The machine intelligence obtained by consumers does not have the technical skills and experience of the excellent intelligent actors they want. The machine intelligence trained by the experience data of multiple different intelligent actors makes the same machine intelligent agent express multiple "split" "robot personalities" without memory communication in the operation, thus showing logical discontinuity, inconsistency, and inconsistent style. Users feel confused and confused, and it is difficult to predict. They feel at a loss as to whether to intervene or the way and degree of intervention. Intelligent actors in different professional fields or time and space regions often experience different behaviors corresponding to different situations. Therefore, the training data generated by the mixture of multiple different intelligent actors (learning objects) often lacks the pertinence corresponding to specific professional fields or specific time and space regions. The industry generally pursues universality excessively when training machine intelligence and fails to take into account pertinence. It blindly and unfoundedly expects that machine intelligence will one day have the general intelligence of drawing inferences from one instance and applying it to other similar cases like humans. The machine intelligence trained in this way often performs poorly in specific professional fields or specific time and space regions, and it is difficult to meet consumers' needs for specialization and localization.The collected training materials often need to be manually screened, classified and labeled to remove materials that do not meet the requirements and leave only materials that meet the requirements, so as to ensure that the imitation behavior generated after machine intelligence training also meets the requirements. Such manual screening, classification and labeling work is often not completed by the intelligent behavior entity itself as the learning object, but is often completed by other entities such as employees or outsourced personnel of the machine intelligence R&D company. The high cost of manual screening, classification and labeling work is generally borne by the machine intelligence R&D entity. In addition, due to the huge workload of screening, classification and labeling and manual operation, it is difficult to check for possible human errors in each training material during quality inspection, and errors in the training material deliverables are inevitably missed. The more serious human errors hidden in the massive training materials will have a significant negative impact on the training results. The entities responsible for the quality of the training materials are divided into three categories: The training materials are scattered and fluid, the producers and quality inspectors of the training materials are separated, the output structure of the training materials is complex, difficult to trace back, and the training quality is uncontrollable; many machine intelligence training materials only contain results, but not the detailed process of reaching the results, which actually weakens the richness of experience and the breadth and depth of intelligent representation information. In addition to the results produced by machine intelligence, many machine intelligence users also hope to obtain or experience the detailed process of producing the results. If machine intelligence only provides users with a result in a simple and rough way, users can only obtain a shallow user experience and can only have a shallow level of surface understanding of the results, "knowing what it is but not why it is so". On the other hand, if machine intelligence only gives results without detailed processes during operation, users cannot intervene in time during its operation, errors are easily accumulated and amplified, and the operation quality is uncontrollable. ;

[0007] It is particularly important that current machine intelligence users generally use machine intelligence trained with other unfamiliar intelligent actors as learning objects, which makes users lack a sense of familiarity in predicting the performance of machine intelligence and unable to make empirical predictions from the beginning. In some machine intelligence usage scenarios, users strongly hope to have another familiar "my intelligent avatar" (rather than "general artificial intelligence") to help them with their work or collaborative work, and hope that machine intelligence "seems to have recreated me". They hope to use capable, familiar, and loyal "like me" intelligent actors rather than unfamiliar general artificial intelligence. Such machine intelligence will help them in a way that they are familiar with and prefer, and they will be particularly tacitly connected, have an organic division of labor, and cooperate tacitly with it, producing a unique effect of "cooperating with themselves". For example, when using smart driving, a smart car user who is an excellent driver hopes that the machine intelligence of his smart car can have "civilized driving habits as good as himself", "rich driving experience as himself", and "familiarity with local traffic conditions as himself". He can easily predict the performance of this machine intelligence based on his own experience and habits, and enjoy the service of "himself" with peace of mind; smart painting users hope that they can have "the style of painting like themselves" and "the style of the painting process like themselves" The machine intelligence can help them establish the same style of draft when they are making drafts, and they can still produce drawing drafts or guidance drawings in the same style when they are not painting, and even after they die. The style process of painting can be simulated and passed down as a teaching example. This example method is not the traditional teaching video that is the same every time, but a simulation of the painting process that is slightly different each time but has exactly the same style. Players of electronic games or online games hope to be able to collaborate and cooperate with a "machine intelligent clone like themselves" who "plays as well as themselves and has a tacit understanding, organic division of labor, and tacit cooperation" in the game. This machine intelligent "player clone" does not play the game according to a mechanical program arrangement, but has flexible freedom. Each time it performs a game task, the details of the process are different, but they all have the same technical skills, experience awareness, style preferences, and rhythm and pace as themselves.

[0008] At present, the complexity of the neural network of machine intelligence trained by humans using neural networks has exceeded the limit of human understanding in a reductionist way. In this case, any efforts to widen and deepen the neural network and expand the neural network model (the so-called "big model") still remain at the current level of complexity that humans can no longer understand. Since humans can no longer understand it, it is no longer possible to continue to increase the complexity to a higher and deeper level. At the same time, the machine intelligence trained in this way is based on past experience. In the process of using stereotyped machine intelligence, new experiences often no longer become new training materials, making stereotyped machine intelligence gradually outdated and increasingly difficult to deal with new situations. Summary of the invention

[0009] The present invention is a personalized, stylized, specialized and localized machine intelligence training method, which uses only the experience created by a single individual intelligent actor as training material, and conducts complete end-to-end training on machine intelligence, so that the neural network machine intelligence can imitate the individual technical skills, experience awareness, style preferences, and rhythm and pace of the intelligent actor.

[0010] As training materials, if the individual experience of intelligent actors focuses on a specific professional field or a specific time and space region, targeted training results will be obtained for the specific professional field or specific time and space region, and the training results will be particularly adapted to or applicable to the specific professional field or specific time and space region. In a specific professional field or specific time and space region, in addition to direct experience, indirect experience can also be added as training materials. In a specific professional field or specific time and space region, intelligent actors can apply machine intelligence trained with themselves as training objects. This application can be one-to-many, that is, intelligent actors simultaneously apply multiple machine intelligences trained with themselves as training objects.

[0011] Machine intelligence can be trained using the experience created by a single individual of an intelligent actor as training material, and then the next machine intelligence can be trained using the experience created by a single individual of the intelligent actor to which the trained machine intelligence belongs. This process can be iteratively extrapolated.

[0012] During training, the machine intelligence can be trained with the real-time experience data of the intelligent actor, or it can be trained with the past experience data records of the intelligent actor.

[0013] The individual of the intelligent agent autonomously selects the experience material used to train the machine intelligence. Experience not only expresses the result, but also emphasizes the process of achieving the result. That is, the experience material as training material is not just a final result, but also focuses on the detailed process of gradually achieving the result step by step.

[0014] The so-called "single individual of an intelligent actor" refers to an intelligent actor that can only express a single behavior at the same time. Its behavior obviously originates from its intelligence, but the source of its intelligence is not required to be unique. For example, an "individual" can make a complex game among multiple ideas after careful consideration and finally make a decision and take a unique action. A "machine intelligent agent" can make a decision and take a unique action after a complex game of decision emergence and priority ranking among multiple intelligent modules or robot characters. A "human-machine community" can make a decision and take a unique action after a complex game of decision emergence and priority ranking among human decisions and machine decisions. That is, at the same point in time, "ideas" can be diverse and "intelligence" can be plural, but "decision" can only be unique, and therefore "behavior" can also be unique.

[0015] The beneficial effects brought about by this patent are described as follows.

[0016] Using only the experience created by a single individual intelligent actor as training material and conducting a completely "end-to-end" training of machine intelligence will enable the neural network machine intelligence to imitate the single individual intelligent actor. The trained machine intelligence will exhibit the technical skills, experience awareness, style preferences, and rhythm and pace of the single intelligent actor. Due to the obvious consistency and identity of the individual's technical skills, experience awareness, style preferences, and rhythm and pace, compared with the mixture of experience data of multiple different intelligent actors, the distribution difference of the experience data created by a single individual intelligent actor is relatively small, and the fitting effect of the training result model is better. The machine intelligence will express the individual style of the specific intelligent actor, so that the individual style will not disappear in the mixture of experience data of multiple different intelligent actors. The individual experience of excellent intelligent actors will be inherited, and consumers can obtain machine intelligence with the technical skills, experience awareness, style preferences, and rhythm and pace of the excellent individual intelligent actors they want, thereby meeting consumers' demand for personalized and stylized machine intelligence. As training materials, if the experience of a single intelligent actor focuses on a specific professional field or a specific time and space area, the trained machine intelligence will be more targeted and perform better for the specific professional field or specific time and space area, which will meet the consumer's demand for the specialization and localization of machine intelligence. Consumers will be able to choose a machine intelligence trained by a specific specialized and localized intelligent actor. The reputation or brand that has been tested or popular will be selected and passed on. In a specific professional field or a specific time and space area, adding indirect experience as training materials in addition to direct experience will increase the amount of data expressing the technical skills, experience awareness, style preferences, and rhythm and pace of the single intelligent actor, making the data dimension richer and the training effect better. The machine intelligence trained only with the experience of a single intelligent actor as training material has natural internal consistency. It will not express multiple "split" "robot personalities" without memory communication, but express a single "robot personality". It will not show logical discontinuity, inconsistency, and inconsistent style. Users expect consistency, logical continuity, and unified style, which is easier to predict and more reliable for whether to intervene or the way and degree of intervention.When users train machine intelligence with themselves as the training object and their own experience as the training material, they will naturally have the opportunity to use this machine intelligence in the future. Other users will also know that they are the training object when using this machine intelligence. The quality of the training materials can only be traced back to themselves, so the subject responsible for the quality of the training materials is single and unique, the producer and quality inspector of the training materials are unified, the output structure of the training materials is simplified, and users are naturally motivated to ensure and improve the quality of the training materials, actively pursue the cleanliness, neatness and purity of the training materials produced by their experience, and actively and independently screen, select and clean up the experience they created, and remove the experience fragments that contain obvious major errors. In addition, most of the time, their training materials are just the raw data of their normal operating experience, and no manual screening and labeling are required. Therefore, for the research and development subjects of machine intelligence, it not only reduces or even eliminates the manual workload of screening, classification, labeling and cleaning of training materials, but also reduces or even eliminates the workload of quality review of training materials, reduces the cost of machine intelligence research and development, and high-quality training materials also make the training quality more controllable. The training materials not only include results, but also emphasize the detailed process of achieving the results, which will increase the richness of experience and the breadth and depth of intelligent representation information. In addition to obtaining the results produced by machine intelligence, users of machine intelligence can also obtain, experience and understand the detailed process of producing the results, thereby obtaining a richer and more complete user experience and having a deep understanding of the results, "knowing why and why". On the other hand, since machine intelligence not only gives results but also detailed processes, it is convenient for users to intervene in the process in a timely manner to avoid the accumulation and amplification of errors, thereby making the quality of work more controllable.

[0017] It is particularly important that when users use machine intelligence trained with their own intelligent behavior as the learning object and their own experience as the training material, they will be familiar with the performance of machine intelligence and have a strong ability to make empirical predictions from the beginning. This machine intelligence with the same technical skills, experience awareness, style preferences, and rhythm as "themselves" will be able to establish a foundation for their work, assist and help, and even let "machine intelligence with the same style as themselves" try first to give themselves some inspiration when they are undecided. In some machine intelligence usage scenarios, users strongly hope to have another familiar "my intelligent avatar" (rather than "general artificial intelligence") to help them with their work or collaborative work, and hope that machine intelligence "seems to have recreated me", and hope to use capable, familiar, and loyal "like me" intelligent behavior rather than unfamiliar general artificial intelligence. Such machine intelligence will help itself in a way that it is familiar with and prefers, and it will be particularly tacit, organically divided, and tacitly coordinated with it, producing a unique effect of "cooperating with oneself". Such needs will be met by the technical solution of this patent. Obviously, the cooperation between people and themselves is the best cooperation. For example, when using smart driving, a smart car user who is an excellent driver can use the machine intelligence of "civilized driving habits as good as his own", "driving experience as rich as his own", and "familiarity with local traffic conditions as himself". When using it, he can easily predict the performance of this machine intelligence based on his familiarity with himself, his own experience and habits, and what specific experience he remembers in the training materials he created, and intervene in a targeted manner without rashness, and enjoy the service provided by "himself" with peace of mind; smart painting users can have and use machine intelligence of "the style of the painting is like his own" and "the style of the painting process is also like his own", which helps him to establish the foundation of the same style of draft when making the draft, and he can still do so when he is not painting. Produce drawing drafts or guidance drawings in the same style, and even after one's death, the painting style process can still be simulated and passed down as a teaching example. This example method is not the traditional teaching video that is the same every time, but a simulation of the painting process that is slightly different each time but has exactly the same style. Players of electronic games or online games can collaborate and cooperate with "machine intelligent clones like themselves" who "play as well as themselves and have a tacit understanding, organic division of labor, and tacit cooperation" to play games. This machine intelligent "player clone" does not play the game according to a mechanical program arrangement, but has flexible freedom. The details of the process are different each time the game task is performed, but they all have the same technical skills, experience awareness, style preferences, and rhythm and pace as themselves.

[0018] Machine intelligence is trained using the experience created by a single individual of an intelligent actor as training material, and then the experience created by a single individual of the machine intelligence body where the trained machine intelligence is located is used as training material to train the next machine intelligence. This process can be iterated and extrapolated. Machine intelligence can understand the complexity of neural networks that humans cannot understand in its own intrinsic way, so it can continue to increase complexity to higher and deeper levels in a hierarchical manner, rather than simply and crudely expanding the neural network model at the low level of the current level of neural network complexity like humans. At the same time, constantly training new machine intelligence with new experiences will enable machine intelligence to keep pace with the times and respond to new situations. DETAILED DESCRIPTION

[0019] The present invention is a personalized, stylized, specialized and localized machine intelligence training method, which belongs to the field of machine intelligence and artificial intelligence. Only the experience created by a single individual of an intelligent actor is used as training material, and the machine intelligence is trained completely end-to-end, so that the neural network machine intelligence imitates the individual intelligent actor. The trained machine intelligence will show the individual technical skills, experience awareness, style preference, rhythm and pace of the intelligent actor, and have natural internal consistency, recreating "my intelligent avatar", and cooperating with it will be particularly tacit, organically divided, and tacitly coordinated. It is more targeted for specific professional fields or specific time and space regions. Consumers can choose the machine intelligence trained by a specific intelligent actor individual, and the long-tested or popular reputation or brand will be selected and passed on. The iterative extrapolation training will enable the complexity of the neural network to enter a higher and deeper level in a hierarchical manner, and can keep pace with the times to cope with new situations.

[0020] One embodiment is the training of intelligent driving machine intelligence. The process of human driving a vehicle is actually a human creation under the guidance of task objectives, traffic rules constraints, and changing environment interactions. Even if the driving tasks are exactly the same, the specific process of completing the same driving task each time is different, the action rhythm may be different, the driving route may be different, and the impact of encountering other traffic participants may also be different. These tasks, actions, routes, states, encounters with other traffic participants, traffic environment and other information can be digitized as training materials for "end-to-end" training of intelligent driving machine intelligence.

[0021] The driver and the vehicle together constitute an intelligent actor. A vehicle with a driver is a single intelligent actor. The intelligent driving machine intelligence is trained end-to-end with its experience data. For example, taxis or online ride-hailing cars in a certain city collect experience data created by a single individual during driving operations to form training materials. Drivers have different style preferences. Some are radical, some are conservative, some are decisive, some are modest, some like to drive faster, and some like to drive slower. The trained machine intelligence will have the individual style embodied by the combination of the driver and his vehicle, and completely imitate the technical skills, experience awareness, style preferences, and rhythm of this single intelligent actor. Therefore, when consumers choose intelligent driving products, they can choose from multiple different machine intelligence styles corresponding to multiple different driver styles. They have room for choice, that is, "you can eat small meals instead of eating big meals", which meets the personalized and stylized user needs, rather than being forced to accept the same general machine intelligence sold by current intelligent driving manufacturers to all their users.

[0022] Currently in the industry, the machine intelligence of smart cars is mostly trained with experience data from multiple different intelligent actors (learning objects), which is often less targeted. Vehicles and their drivers traveling in different cities, regions, and areas deal with different landforms (for example, some cities are plains, and some cities are mountainous), different forms of transportation infrastructure (for example, despite the constraints of national standards, traffic lights in different cities may have different designs and practices in terms of lamp heads, lamp posts, lampshades, column decorations, spacing, installation angles, etc.), different styles of transportation infrastructure (for example, some cities have many viaducts, some cities have railings on the center line of the road, and some cities have mixed traffic of people, vehicles, and livestock on the streets), and different climatic environments (for example, some cities have rain all year round, some cities often have sandstorms, and some cities have no wind or traffic). There are often smog in some cities), different lighting conditions (some cities have strong sunny days all year round, while some cities have insufficient light due to cloudy days all year round), so, the empirical data collected from multiple cities, regions and areas are mixed together to train the machine intelligence. Obviously, it is difficult to take into account the specificity while pursuing universality. The designers of machine intelligence hope that their smart cars can automatically drive in these different cities, regions and areas like humans, but the result is that when targeting a specific city, region or area, they often show insufficient performance and poor response, which makes it difficult to meet consumers' demand for specialization and localization. In this technical solution, in most cases, the operating scope of taxis or online car-hailing services is generally in the same city, region, or area. Therefore, the experience data they create is often more targeted at this city, region, or area. The adoption of this technical solution will obviously further improve the pertinence to this city, region, or area, and enhance the targeted performance of machine intelligence trained with such experience data in this city, region, or area. For example, among the taxis or online car-hailing services in a certain city, a number of "civilized taxi" drivers and a number of "civilized online car-hailing" drivers are selected by the organization, and vehicles with the same performance specifications (such as chassis specifications and power indicators) are used for driving operations and experience data is collected. Each "civilized taxi" driver and "civilized online car-hailing" driver A machine intelligence will be trained end-to-end for each "Ming online car-hailing" driver. After a period of one-on-one special training, the machine intelligence will be able to imitate the driver's technical skills, experience, style preferences, and rhythm and pace better. Then these intelligent driving machine intelligences can be sold. When sold, they can be marked with labels such as "Full-time training by XXX, winner of the title of civilized taxi in a certain city in a certain year", "Full-time training by XXX, taxi driver who operated safely without accidents for ** days", "Full-time training by XXX, driver of the March 8th Red Flag Bearer", and "Full-time training by XXX, driver of the online car-hailing car with the best consumer evaluation in a certain city in a certain year" to meet consumers' personalized and stylized selection needs for intelligent driving machine intelligence.For consumers, the experience of using smart driving is "just like the best local taxi or online car-hailing driver X is driving for me". For example, users will experience the specific civilized driving habits of the civilized taxi driver X: in front of the zebra crossing where people are crossing the street, the car will brake a little earlier, 1 meter away from the zebra crossing, so as not to scare the pedestrians because of the close distance. In this way, the reputation or brand that has been tested or popular will be selected and passed on.

[0023] When taxi or online car-hailing drivers create experience data during the driving process, they can also use vehicles with existing intelligent driving foundations to drive in a human-machine co-driving mode, that is, the intelligent car drives automatically in normal times, and the vehicle behavior is driven by the data output by the intelligent car. When the driver intervenes and takes over due to his own needs and style, the vehicle behavior is driven by the data output by the driver. At this time, the data output by the driver overwrites the data output by the intelligent car in the command chain. For human-machine co-driving, regardless of whether the immediate decision is made by human intelligence or machine intelligence, the final driving behavior is unique, and the collected experience data training materials are also unique, including the above-mentioned collection of behavior data output by the intelligent car when it does not intervene and take over and the collection of behavior data output by the driver when it intervenes and takes over. This data set still represents the driver's technical skills, experience awareness, style preferences, and rhythm and pace with assistance, that is, a single intelligent car in human-machine co-driving also belongs to the "single individual of intelligent behavior body" described in this technical solution.

[0024] When the training module and the trained model are on the vehicle, the model is trained in real time with the vehicle's real-time environment and action data during effective driving (operating state, such as normal driving, stopping and waiting for a red light for dozens of seconds). When the vehicle is in an ineffective driving process (stop state, such as the driver stops to rest for 1 hour for lunch, or stops to rest for 10 hours when the driver returns home at night), the training module uses the stored data records for training.

[0025] When the training module and the trained model are on the cloud, the vehicle uploads real-time experience data to the cloud during effective driving (operating state, such as normal driving, stopping and waiting for a red light for dozens of seconds), and the cloud training module trains the model in real time. When the vehicle is not in effective driving (non-operating state, such as the driver stops for a 1-hour rest for lunch, or the driver stops for a 10-hour rest after returning home at night), the cloud training module uses the data records stored on the cloud for training.

[0026] In a specific professional field or a specific time and space area, in addition to direct experience, indirect experience can also be added as training materials. For intelligent driving machine intelligence, the direct experience data in the end-to-end training materials obviously includes task information, state information, environmental information and driving action information. For example, the driver's specific environmental cognition when driving directly leads to specific driving actions under the interaction of task purpose orientation, traffic rules constraints, and changing environment. In addition to these direct experiences, indirect experience data can also be collected, such as collecting GPS geographic positioning latitude and longitude data when the vehicle is driving as training materials. The reason why it is called "indirect" is that the driver never consciously perceives and considers the latitude and longitude data of his location when driving, but because this vehicle operates and collects experience data in a specific time and space area (for example, limited to a certain city) in a targeted manner, its driving behavior and longitude and latitude data can often have a factual approximate match and have indirect correlation. For example, only straight driving or right turning actions have always occurred at a certain longitude and latitude position, and driving from a certain longitude and latitude to a certain longitude and longitude has only occurred in the forward direction but not in the reverse direction. Adding indirect experience as training material will increase the amount of data that expresses the technical skills, experience awareness, style preferences, and rhythm and pace of a single individual of the intelligent actor, making the data dimension richer and the training effect better.

[0027] In the current real-life case, a certain brand of smart car intelligent driving system FSD v12, which uses fully end-to-end training machine intelligence, once had the following real performance: this smart car was queuing in front of a traffic light at a certain intersection, waiting to cross the intersection. The straight red light turned green, and the car in front crossed the line and went straight forward. This car followed and then slowed down and stopped. When the straight traffic light was green, it stopped within the line. After that, the straight traffic light turned from green to red, and then the left turn traffic light turned from red to green. The car in the straight lane and under the control of the straight traffic light (red light) actually started to accelerate. Fortunately, it was stopped by the terrified driver's emergency brake, but at this time the vehicle had crossed the line and violated traffic rules. After technical investigation, it was found that the smart car was originally going to follow the car in front and go straight through the line, but its machine intelligence sensed that all lanes in the road ahead opposite the intersection were full of vehicles and seriously congested. The machine intelligence of this smart car has been trained to learn the driving behavior of many "very gentlemanly" car owners, so at this time it imitated and "also very gentlemanly" did not follow the car in front and go straight through the intersection when the light was green, entering the opposite road to aggravate congestion or occupy the intersection space. For convenience, we call it "Robot Cell 1" here. Afterwards, the straight light changed from green to red, and a few seconds later, the left turn light changed from red to green At this time, the smart car imitates the following human behavior based on the driving behavior of many "normal" car owners that have been trained and learned: "The left turn light is red, the straight light turns from red to green, the car in front crosses the line and goes straight, and this car starts to follow and stops in front of the line. This means that this car is waiting for the left turn light to turn green, that is, when encountering a green light, it does not cross the line, which means that it is waiting for another light to turn green." This is a common and normal human driving behavior. The machine intelligence is "neurotically disordered" and "split personality" connected to this imitation behavior. For convenience, we call it "Robot Personality 2" here. Obviously, in this real case, although the same machine intelligence is performing intelligent driving, "Robot Character 1" and "Robot Character 2" are split, learning to imitate vehicles driven by different people respectively. The two "Robot Characters" do not communicate with each other, and multiple different "Robot Characters" emerge on the spot to imitate human driving experience. However, this kind of imitation without memory communication, discontinuous logic, inconsistency, and inconsistent style manifests itself as confused and clumsy intelligent driving behavior, resulting in red light violations. At the same time, precisely because the performance of its machine intelligence when "Robot Character 1" emerges is relatively ideal, the user is completely unable to anticipate the low-level illegal actions that lack internal consistency when "Robot Character 2" emerges immediately afterwards, and is therefore unable to intervene predictively.Obviously, in this case, if we only use the experience created by a single "vehicle driven by an extremely gentlemanly driver" as training material (the same straight-moving car must be consistent and logically continuous after "gentlemanly" stopping and waiting, and continue to move straight after the road ahead is no longer congested), and conduct complete end-to-end training on the intelligent driving machine intelligence, the neural network machine intelligence will have the internal consistency to imitate the single individual of the intelligent behavior entity, and the trained intelligent driving machine intelligence will show the consistent and continuous technical skills, experience awareness, style preferences, rhythm and pace of this "vehicle driven by an extremely gentlemanly driver", and a single "robot personality" will emerge. There will no longer be memoryless communication, logical discontinuity, inconsistency, and inconsistent style. Users will no longer experience the "unexpected" and "inexplicable" performance of the vehicle in this case, but will be more likely to predict its internally consistent performance, and will have more basis for whether to intervene or the method and degree of intervention.

[0028] When drivers train intelligent driving machine intelligence by using themselves as training objects and their own experience as training materials, the subject responsible for the quality of training materials is single and unique, the producer and quality inspector of training materials are unified, and the output structure of training materials is simplified. Considering that they will naturally have the opportunity to use this machine intelligence in the future, and that other users will know that their training objects are themselves when using this machine intelligence, the quality of training materials can only be traced back to themselves. Drivers will naturally have the motivation to ensure and improve the quality of training materials, actively pursue the cleanliness, neatness and purity of training materials generated by their experience, and will actively and independently screen, classify, select and clean up the experience they have created. Remove experience fragments that contain obvious major errors. For example, if the driver accidentally runs a red light and violates traffic rules during operation and collection of experience data, this is a major mistake. The driver will obviously not want to use it as training material to mislead machine intelligence training. Therefore, the driver actively and autonomously deletes the experience fragments collected several minutes before the violation through relevant operations; when the driver transports a critically ill patient to the hospital for emergency treatment, he deliberately runs a red light continuously. This is a special behavior that is not universal in extremely low-probability events. The driver will obviously not want to use it as training material to mislead machine intelligence training. Therefore, the driver actively and autonomously deletes the experience fragments related to the event through relevant operations. In addition, most of the time, its training materials are just the original data of its normal operating experience, and no manual screening, classification and labeling are required. Therefore, for the research and development subject of intelligent driving machine intelligence, it not only reduces or even eliminates the manual workload of screening, classification, labeling and cleaning of training materials, but also reduces or even eliminates the workload of quality review of training materials. The research and development subject of intelligent driving machine intelligence reduces costs, and high-quality training materials also make the training quality more controllable. In this technological and economic structure formed, the machine intelligence generated by individual driver training and cultivation, their individual technical skills, experience awareness, style preferences, and rhythm and pace form a time-tested or popular reputation or brand that is designated, selected and passed on by consumers. Consumers can pay to purchase from this driver (the accumulator and provider of training materials), and smart driving manufacturers and drivers are also jointly responsible or compensated for any problems that arise. Consumers, smart driving manufacturers and individual drivers (accumulators and providers of training materials) can share the benefits and risks of smart driving.

[0029] The training of machine intelligence can keep pace with the times. For example, driver Li Si buys an intelligent driving vehicle that has been "trained by Zhang San, the winner of the title of civilized taxi", and operates a taxi. The experience data created in the operation is used to train the next (next generation) new intelligent driving machine intelligence. After a period of time, the taxi operated by Li Si is also selected to be "civilized taxi". Then this new generation of intelligent driving machine intelligence can be put on sale, and the label "from a certain year to a certain year, Zhang San, the winner of the title of civilized taxi, trained by a special person, and from a certain year to a certain year, Li Si, the winner of the title of civilized taxi, continued to train and iterate" is marked when selling, meeting the personalized and stylized selection needs of consumers. For consumers, their experience of intelligent driving is "the best local taxi drivers Zhang San and Li Si continue to serve me, their training time is very long, so they are very experienced, and they have been training until recently, so their experience is very fresh and can cope with new situations". In this way, the individual's time-tested or popular reputation or brand will be selected and passed on. Machine intelligence can understand the complexity of neural networks that humans cannot understand in its own intrinsic way, and can train new machine intelligence with the experience created by the intelligent actors in which machine intelligence resides. It can continue to increase the complexity to higher and deeper levels in a hierarchical leap, rather than simply and crudely expanding the neural network model at the current low level of neural network complexity like humans. Continuously training new machine intelligence with new experiences will enable machine intelligence to keep pace with the times and respond to new situations. For example, during Li Si's operational training, some new traffic sign infrastructure appeared in the urban area. These new situations did not appear during Zhang San's operational training. The "old" smart driving vehicles "trained by Zhang San, the winner of the title of civilized taxi, throughout the whole process" will be unable to cope with it. However, the new generation of smart driving vehicles "trained by Zhang San, the winner of the title of civilized taxi, throughout the whole process from a certain month to a certain month of a certain year, and continued to be trained and iterated by Li Si, the winner of the title of civilized taxi, from a certain month to a certain month of a certain year" will have the ability to cope with it.

[0030] For a winner of a certain civilized taxi title in a certain city, this driver can use a machine intelligent vehicle trained with himself as the training object and his own experience as the training material to perform intelligent driving operations. When using such intelligent driving, excellent drivers will experience "civilized driving habits as good as themselves", "driving experience as rich as themselves", "familiarity with local traffic conditions as themselves" machine intelligence, and enjoy "myself" serving them with peace of mind. Because this machine intelligent vehicle is a "my smart driving clone", it is familiar with its own technical skills, experience, style preferences, and rhythm and pace. Therefore, the driver will have a particularly strong ability to predict, and have a high degree of self-confidence and operational execution ability in what situations need to intervene and take over, and what situations do not need to intervene and take over. The more familiar the places, scenes, and actions are, the less intervention is needed, and the more rare the places, scenes, and actions are, the more special attention and vigilance are needed. When using their own machine intelligence "intelligent driving clone" smart car, the driver can supervise and control it inside the car, or remotely supervise and control it outside the car. Because they are particularly familiar, at ease, and confident, they can use, monitor, and manage multiple "intelligent driving clone" smart cars of their own at the same time. When the driver is relatively old, he can use the machine intelligence "intelligent driving clone" trained by himself, and no longer need to sit in the car and drive only one taxi to operate the business. Instead, he can remotely monitor and manage multiple "intelligent driving clone" smart cars of his own with high reliability in a comfortable room sheltered from wind and rain. He can rely on his lifelong local experience, with high confidence and high probability to successfully predict which situations the smart car can reasonably respond to and which special situations may make it difficult for the smart car to cope with, because he is familiar with the local street environment and understands what experience he used as training material to train his smart car machine intelligence. Your own "smart driving clone" will help you in a way that you are familiar with and prefer, and you will be particularly in tune with it, with an organic division of labor and tacit cooperation. For example, one day a street encounters a float parade. The normal traffic environment changes suddenly, but the road structure of the street itself does not change. At this time, the driver has a lot of experience driving on this street and is confident that the smart car he has trained with experience can turn highly reliably. He is just worried about how the smart car will deal with the extremely complex traffic participants (floats, pedestrians, dancers, waving colorful flags, flying paper ribbons and balloons, etc.) in the sudden float parade (worrying that the smart car will brake too frequently and too cautiously). In this case, the driver can instruct the smart car's machine intelligence to be responsible for steering, while he or she will temporarily be responsible for the vehicle's throttle and brakes, which can safely avoid non-collisionable traffic participants, and will not be forced to stagnate for a long time by pseudo-obstacles such as essentially harmless ribbons and balloons, forming such a tacit and organic division of labor and cooperation. Obviously, the cooperation between people and oneself is the best cooperation.

[0031] One embodiment is the training of intelligent painting machine intelligence. The process of human painting is actually a kind of human creation guided by the purpose of expression, constrained by tool rules, and restricted by changing environments. Even if the painting expression task is exactly the same, the specific process of completing the same task each time is different. The rhythm of the pen action may be different, the specific painting process may be different, and the impact of the painting restriction environment may also be different. These tasks, actions, processes, environments and other information can be digitized as training materials for end-to-end training of intelligent painting machine intelligence.

[0032] Only the experience data of a certain painting master's personal creative process is used as training material, and machine intelligence is trained in a completely end-to-end manner. For example, video images of the action sequence and the picture area reflecting the entire painting process of a certain painting master are collected. These materials will reflect the painting master's technical skills, experience awareness, style preferences, rhythm and pace, including the composition of the entire picture, the choice of pictures and blank spaces, the sequence and thickness of each stroke, the choice and matching of colors, the weight and change of pen force, etc. The trained intelligent painting machine intelligence will have the ability to imitate the master's superb technical skills, experience awareness, style preferences, and rhythm and pace. Since the training materials come from one person, the distribution difference of the experience data is relatively small, and the fitting effect of the training result model is better. Consumers can choose the painting machine intelligence trained with the painting master as the learning object to meet consumers' demand for personalized and stylized machine intelligence. A well-tested or popular reputation or brand will be selected and passed on.

[0033] A master painter who specializes in a certain professional field can be used as a learning object, such as a master painter who specializes in Chinese flower and bird paintings. Then the trained machine intelligence will also have the same superb technical skills, experience awareness, style preferences, and rhythm and pace in the professional field of flower and bird paintings as this master, and perform better in this professional field of flower and bird paintings. Consumers can choose the painting machine intelligence trained by the master painter of flower and bird paintings as a learning object to meet consumers' needs for the specialization and localization of machine intelligence. The reputation or brand that has been tested or popular will be selected and passed on. For painting, in addition to the direct experience of the composition of the entire picture, the choice of pictures and blank spaces, the order and thickness of each stroke, the choice and matching of colors, the weight and change of pen strength, there are actually indirect experiences in the experience created by the painting master, such as the thickness of rice paper, the viscosity of pigments, the size of gold powder particles used to outline the petals and bird tails, etc. These indirect experiences can be quantified into data and entered into the training materials. The amount of targeted training data in professional fields is increased, the data dimension is richer, and the training effect is better.

[0034] When the painting master trains machine intelligence using himself as the training object and his own experience as the training material, the subject responsible for the quality of the training materials is single and unique, the producer and quality inspector of the training materials are unified, and the output structure of the training materials is simplified. Considering that he will naturally have the opportunity to use this machine intelligence in the future, and considering that other users will know that their training object is themselves when using the machine intelligence, the traceability of the quality of the training materials can only point to himself. He will naturally have the motivation to ensure and improve the quality of the training materials, and actively pursue the cleanliness, neatness and purity of the training materials produced by his experience. He will actively and autonomously screen, classify, select and clean up the experience he created, and remove experience fragments that contain obviously major errors. For example, major mistakes occasionally occur during the painting process, which may have a major adverse impact on the training results. The painting master will actively and autonomously delete this experience data. In addition, most of the time, the training materials are just the raw data of their normal painting experience, which does not require manual screening and labeling. Therefore, for the research and development entities of machine intelligence, it not only reduces or even eliminates the manual workload of screening, classifying, labeling, and cleaning training materials, but also reduces or even eliminates the workload of quality review of training materials. The research and development entities of machine intelligence reduce costs, and high-quality training materials also make training quality more controllable.

[0035] A painting master can use machine intelligence trained with himself as the training object, and can have and use machine intelligence with "painting style like himself" and "painting process style like himself" to help himself establish the same style draft foundation when making a draft, and can still produce the same style of drawing drafts or guidance drawings when he is not painting. Even after his death, he can still pass on the painting style process as a teaching example. This example method is not a traditional teaching video that is the same every time, but a painting process simulation that is slightly different but completely consistent in style every time. When he is hesitant about the next stroke in the painting process, he can let machine intelligence like himself try to draw a stroke first to give himself some inspiration. If a painting master needs to mass-produce paintings, he can monitor and manage multiple machine intelligences trained with himself as the training object to perform painting operations at the same time, and use his own experience to monitor and guide the painting process and screen and improve the painting results.

[0036] The experience created by a single individual of a machine intelligence trained with a master painter as the training object can be used as training material to train the next (next generation) machine intelligence, and this process can be iterated and extrapolated. Machine intelligence can understand the complexity of neural networks that humans cannot understand in its own intrinsic way, so it can continue to increase complexity to higher and deeper levels in a hierarchical manner, rather than simply and crudely expanding the neural network model at the current level of neural network complexity like humans. At the same time, continuous training of new machine intelligence with new experiences will enable machine intelligence to keep pace with the times and respond to new situations, such as adapting to and training for a new type of pigment that has just emerged.

[0037] Machine intelligence can be trained using the experience data of a master painter's creative process while he is painting, or it can be trained using the experience data of the creative process recorded in the past.

[0038] The training materials formed by the painting experience of painting masters not only include the results, but also emphasize the detailed process of achieving the results, that is, they pay special attention to collecting the entire process of painting, which will increase the richness of painting experience and the breadth and depth of painting intelligence representation information. In addition to obtaining the painting results produced by machine intelligence, intelligent painting users can also obtain, experience and understand the detailed process of the output results, such as every stroke of the painting process, what color is used to cover what color, and how the specific outline is obtained, so as to obtain a richer and more complete user experience and have a deep understanding of the results of "knowing what it is and why it is". On the other hand, since such painting machine intelligence not only gives the painting results but also gives the detailed process, it is convenient for users to intervene in time during the machine intelligent painting process, such as immediately stopping and updating the instructions when a certain stroke obviously deviates from the user's expectations, so as to avoid the accumulation and amplification of errors, thereby making the quality of painting operations more controllable.

[0039] One example is the training of machine intelligence for "player avatars" in electronic games or online games. The process of humans playing games is actually a human creation guided by task objectives, constrained by game rules, and restricted by changing environments. Even if the game tasks are exactly the same, the specific process of completing the same task each time is different. The action rhythm may be different, the skill selection may be different, the route may be different, and the impact of encountering other game players and environmental obstacles may also be different. These tasks, actions, skills, routes, encounters with other players, environmental obstacles and other information can be digitized to serve as training materials for end-to-end training of machine intelligence game players.

[0040] Only the gaming experience created by a certain game player is used as training material to train machine intelligence in a completely end-to-end manner. For example, information data such as tasks, actions, skills, routes, encounters with other players, and environmental obstacles during the entire game process of a famous player can be collected. This game player is a game star or a game master who is generally loved by game players. The trained gaming machine intelligence will have the ability to imitate the superb technical skills, experience awareness, style preferences, and rhythm and pace of this famous player. Since the training materials come from one person, the distribution differences of the experience data are relatively small, and the fitting effect of the training result model is better. Consumers can choose a machine intelligence player avatar trained with the famous player as the learning object, and get auxiliary guidance from a game star or master when they play the game themselves, satisfying consumers' demand for personalized and stylized machine intelligence. A time-tested or popular reputation or brand will be selected and passed on.

[0041] A famous player may focus on a fixed game, or a fixed role in a game, or a fixed task in a fixed map version of a game. The machine intelligent player avatar trained with the experience collected in this way will be particularly focused on these local areas, showing both better game performance and the personal preferences of the famous player, satisfying consumers' demand for specialization and localization of machine intelligence.

[0042] When a famous player trains machine intelligence using himself as the training object and his own experience as the training material, the entity responsible for the quality of the training materials is single and unique, the producer and quality inspector of the training materials are unified, and the output structure of the training materials is simplified. Considering that he will naturally have the opportunity to use this machine intelligence in the future, and considering that other users will know that their training object is themselves when using this machine intelligence, the traceability of the quality of the training materials can only point to himself. He will naturally have the motivation to ensure and improve the quality of the training materials, and actively pursue the cleanliness, neatness and purity of the training materials produced by his experience. He will actively and independently screen, classify, select and clean up the experience he created, and remove experience fragments that contain obviously major errors. For example, major mistakes such as accidental injury to friendly forces may occasionally occur during the game, which may have a major adverse impact on the training results. This famous player will actively and independently delete this experience data. In addition, most of the time, the training materials are just raw data from their normal gaming experience, which does not require manual screening, classification, and labeling. Therefore, for the research and development entities of machine intelligence, it not only reduces or even eliminates the manual workload of screening, classifying, labeling, and cleaning up training materials, but also reduces or even eliminates the workload of quality review of training materials. The research and development entities of machine intelligence reduce costs, and high-quality training materials also make training quality more controllable.

[0043] Game players can use machine intelligence trained with themselves as the training object, and cooperate with "machine intelligence clones like themselves" in the game, which "plays as well as themselves and has a tacit understanding, organic division of labor, and tacit cooperation". This machine intelligence player clone does not play the game according to mechanical programs, but has flexible freedom. The details of the process are different each time the game is performed, but they all have the same technical skills, experience, style preferences, and rhythm as themselves. Under permitted conditions, a player may use one or more "machine intelligent clones like himself" to play the game without personally operating the game. Many online games do not allow players to use ordinary program robots with fixed routines to play the game without personally operating the game. Such ordinary program robots perform game operations completely according to fixed programs. Online game operators can screen out players who output repetitive, programmed and patterned game operations, and then automatically eject them from the game server. However, if the "machine intelligent clone like yourself" trained using the present technical solution, even if the game tasks are exactly the same, the specific process of each "game clone" completing the same task is different, the action rhythm may be different, the skill selection may be different, the travel route may be different, and the corresponding reactions when encountering other game players and environmental obstacles are also different. It has the game technical skills, experience awareness, style preferences, and rhythm pace that this human player has always had, so it will not be identified as a program robot by the game operator server and will not be ejected from the game server. Users can also "play games with themselves", that is, team up with one or more "machine intelligent avatars like themselves" to play multi-player games, or use "machine intelligent avatars like themselves" to operate movement and positioning in single-player games while concentrating on manually manipulating attack movements or casting skills. Such organic division of labor and cooperation can enable individual players to take into account both ideal displacement (such as "body movement") and action (such as "attack and defense") in many complex scenarios. That is, there is no need to explicitly instruct the game avatar how to move, because you know how it will run (just like you know how you will run), and you can focus all your attention on manually manipulating game attack movements or casting game skills, without having to consider positioning and affecting your concentration. This is a kind of organic division of labor, cooperation and coordination. Obviously, the cooperation between people and themselves is the best cooperation and coordination.

[0044] The game experience created by a certain game player can be used to train machine intelligence players, and the experience created by this machine intelligence player can then train the next (next generation) machine intelligence player. This process can be iterated and extrapolated continuously. Machine intelligence can understand the complexity of neural networks that humans cannot understand in its own intrinsic way, so it can continue to increase the complexity to higher and deeper levels in a hierarchical manner, rather than simply and crudely expanding the neural network model at the current low level of neural network complexity like humans. At the same time, constantly training new game machine intelligence with new experience will enable game machine intelligence to keep pace with the times and respond to new situations, such as adapting to and training for new tasks, new roles, and new gameplay that appear in new versions of the game.

[0045] Game players can conduct real-time training while creating game experience data, or they can store their experience data for training.

Claims

1. A personalized, stylized, specialized and localized machine intelligence training method, characterized by: The machine intelligence is trained completely end-to-end using only the experience created by a single individual intelligent actor as training material.

2. The method according to claim 1, characterized in that: As training material, the individual experience of intelligent actors focuses on specific professional fields or specific time and space regions.

3. The method according to claim 2, characterized in that: In a specific professional field or a specific time and space area, in addition to direct experience, indirect experience is added as training material.

4. The method according to claim 2, characterized in that: Intelligent actors use machine intelligence trained on themselves.

5. The method according to claim 4, characterized in that: Intelligent actors apply machine intelligence trained on themselves in a one-to-many manner.

6. The method according to claim 1, characterized in that: Machine intelligence is trained using the experience created by a single individual of an intelligent actor as training material, and then the next machine intelligence is trained using the experience created by a single individual of the intelligent actor to which the trained machine intelligence belongs as training material, and this process is iteratively extrapolated.

7. The method according to claim 1, characterized in that: Train machine intelligence with real-time experience data of intelligent actors.

8. The method according to claim 1, characterized in that: Machine intelligence is trained with past experience data records of intelligent actors.

9. The method according to claim 1, characterized in that: A single individual intelligent agent autonomously selects the empirical material used to train machine intelligence.

10. The method according to claim 1, characterized in that: Experience not only expresses the results, but also emphasizes the process of achieving the results.