Fundamentals of the approach to the construction of universal intelligence. Part 2
Part 1.
When creating a strong AI, it is natural to reproduce, if not all the details of the work of the human brain, then at least those functions that it performs. Otherwise, it is very difficult to be sure that it is the intellect that is created. It is this goal that cognitive architectures pursue, which combine functions such as learning, memory, planning, etc., that is, all (or almost all) of what is in natural intelligence. This makes cognitive architectures so attractive and popular.
However, in itself, the desire to endow the computer with all the same cognitive functions that a person has does not say how to do it correctly. As a result, a lot of cognitive architectures have been developed to date, a number of which are often positioned as a way to build a strong AI. These include, in particular, such popular architectures as “builders of strong AI”, such as Soar and ACT-R.
Many architectures often start from the phenomenology of higher cognitive functions of the human mind. However, due to the lack of a complete understanding of the nature of these functions and the requirements for them, their implementation is largely arbitrary.
Often, even the construction of such architectures is carried out within the framework of the traditional symbolic approach that models only the “tip of the iceberg” of human thinking. Nevertheless, an attempt is often made to construct architectures that reproduce not only high-level, but also low-level functions (so-called emergent architectures). Moreover, AI researchers are well aware of the need to combine symbolic and subsymbol levels and the development of hybrid architectures, as well as the need to build embodied systems (which are key, in particular, to obtain the semantic basis of concepts), which are highly problematic in purely symbolic architectures (see [ Duch et al., 2008] as a review).
Nevertheless, it is noted [Duch et al., 2008], which is very rarely possible to use for solving real problems, not to mention scaling to the level of autonomous behavior in the real world. So why did cognitive architectures not lead to significant progress in the field of strong AI? The answer to this question has already been given above.
These systems are most likely doomed to non-universality, since they are assembled from weak components. This apparently also applies to such systems that were initially positioned as general intelligence systems, such as Novamente (described in [Goertzel and Pennachin, 2007]). Of course, the possibility of introducing the property of universality as an extension of a particular architecture is not excluded (after all, the universality of intelligence can hardly be ascribed to most animals, which means it appeared as an evolutionary superstructure over more particular forms of intelligence). Nevertheless, this way seems to us more laborious and less optimal.
This approach is based on the following definition given by P. Wang [Wang, 2007]:
Whereas adaptation (as the ability to learn from experience) is a fairly common requirement, while the main features of the approach are derived from a lack of resources and knowledge (because when resources are and there is enough information, not quite intellectual methods can be used). As a result, within this approach, a categorical logic variant is built to take into account the fuzziness of knowledge, and a distributed knowledge manipulation system is proposed that takes into account the limited computational resources.
In this case, the author proposes to separate the concepts of "intellectual" and "effectively intellectual." Such a division seems to be quite fair and reflects the intuitive impression that, for example, a chess program operating by the “brute force” method is not intellectual in the same sense as the intellectual chess player.
Although it is possible to agree with the principle of effective intelligence, this particular approach can hardly become the basis for building FIS: it omits those aspects of intelligence that are revealed in universal algorithmic models and in cognitive architectures. In other words, the thesis about the need for resource constraints does not say how to enter them correctly.
In particular, this is evident from the fact that P. Wang introduced as a fundamental principle also the lack of knowledge of the agent. The lack of knowledge, of course, is important, but it is fully taken into account in (criticized by Wang) models of universal algorithmic intelligence, which include not only a search in the action space, but also a universal inductive conclusion, for which taking into account the fuzziness of knowledge is not a fundamental principle, but only heuristics to simplify sorting models (which will be demonstrated later).
As a result, within the framework of this approach, only a private cognitive architecture has been developed, which does not have any fundamental advantages over others, although systematically following the principles of limited resources also has considerable heuristic power.
')
The very idea of ​​this approach has been known for a long time, but it received recognition relatively recently mainly through the work of [Hutter, 2001], [Schmidhuber, 2003] and other works of these authors. Within its framework, the focus is on the Solomon’s model of universal induction, included in the system of choice of actions in the environment to maximize a certain estimating function.
Here, the analysis begins with a simple universal model, on which no resource constraints are imposed. The first step of our approach is similar, since we believe that the universality property is highly desirable to immediately introduce into the model of a universal AI and to maintain this property during the development of the model, which is carried out by introducing resource constraints.
In modern versions of the approaches under consideration, resource constraints are also introduced, but with preservation of the maximum impartiality of the universal AI, which allows us to construct general models of self-optimization.
This consideration of resource constraints, however, is not quite sufficient. It can be said that it requires the reproduction of the whole evolution, which also began as a universal self-optimizing search without any a priori information. It is obvious that the formation of such universal intelligence could be accomplished in the foreseeable time, it is necessary to lay in it as a sufficiently large amount of a priori information about the structure of the external world, and heuristics to reduce the search for models and actions. These heuristics can just be gleaned from the phenomenology of the cognitive functions of natural intelligence. On the other hand, it is not rational to manually build too much specific knowledge in a strong AI that he can learn on his own (which is what projects such as Cyc sin for). Obviously, it is necessary to achieve an optimal compromise between these two extremes.
In addition, a separate issue for discussion is whether or not the presented models are universal. For this, it is necessary to carefully compare the hypothetical capabilities of these models with the capabilities of a person. In part, such comparisons are made (for example, [Hutter, 2005]), although they cannot be called indisputable or exhaustive. Nevertheless, doubts about the true universality of these models can be put forward, which will be shown in the analysis of our own model of universal algorithmic intelligence.
Now we note only one of these doubts, which is that intelligence only in the zero approximation can be reduced to maximizing an a priori given objective function. After all, if, say, the task of intelligence is to ensure survival, then an a priori given objective function (based, say, on emotional evaluations) can only be a rough heuristic approximation of the goal of survival. This means the necessity of the existence of special mechanisms that allow in some way to refine the objective function in ontogenesis. Here you can make the following analogy with chess. Let a single intelligent agent play only one game. Having limited computing resources, he cannot perform a full search of options to predict victory or defeat. Being born with a minimum of a priori knowledge of the world, it cannot have a complex objective function that would effectively cut off unpromising options on the game tree. The initial objective function can rely only on some directly perceived stimuli, say the total force of the figures (giving a feeling of pain and pleasure when you lose your figure or eat an opponent's figure). In the process of growing up (a game), an agent can build more complex concepts, but on his own (without having lived life entirely), he cannot, in principle, determine how the objective function can be improved based on these concepts. This information, however, can be provided by other agents, but only if there is some good mechanism for modifying the objective function. This aspect is also related to the problem of friendly AI ...
The problem of learning objective functions is sometimes considered as fundamental in building strong AI (or, more precisely, friendly AI [Yudkowsky, 2011]). Within the framework of this approach, it is rightly noted that maximizing an a priori objective function is insufficient for artificial intelligence to be universal, especially in terms of effective (and desired) interaction with the social environment, which is the same element of objective reality as the physical environment.
The problem of endowing AI with the ability to modify its own objective function is non-trivial due to the fact that it is not clear how the objective function can be optimized, if not under the control of another objective function (or some other a priori mechanisms). The importance of the possibility of modifying an objective function is connected not only with the fact that it is necessary for the full versatility of the agent, but also with the fact that the AI ​​seeking to maximize the a priori objective function may well find such actions optimal from the point of view of this function that will be extremely undesirable for people [Yudkowsky, 2011]. Although the importance of these aspects is indisputable, considering them outside of specific models of universal intelligence does not allow one to chart a way to create a strong AI (or rather, it sets some limitations on the way it is created), therefore this approach should be considered complementary to other approaches. The possibility of modifying the objective function must be provided in the architecture of the universal intelligent agent, although in general this can be considered at the same level as other cognitive functions, namely, as a specific heuristic of increasing the development efficiency of the "infant" AI to the level of "adult" AI.
There is a large line of research in the field of strong AI, associated with the bionic approach. It highlights attempts (see, for example, [Garis, 2007] [Red'ko, 2007]) to model the brain at different levels of detail, to reproduce adaptive behavior, starting from its simplest forms to more complex ones, modeling evolution, self-organization as a whole. Often this approach is imitative in nature and rather rigidly opposed to the algorithmic approach, which is not deep enough. In particular, various simulation models of evolution and self-organization do not lead to unlimited development for the simple reason that their authors do not even try to consider issues related to the computational complexity of the optimization problems to be solved and the algorithmic completeness of those forms of behavior that can in principle be obtained during this modeling. Because of this, it is highly doubtful that the bionic approach itself can lead to the creation of a strong AI. However, at the same time, it can be an important source of productive ideas, which would be too wasteful to ignore.
As you can see, different existing approaches to strong AI do not so much contradict each other, as they consider different aspects of the problem of universal AI, and therefore it is necessary to combine them. Naturally, there are many integration approaches that attempt to synthesize various existing systems and methods, so the idea of ​​integration is generally not new. However, this integration is often limited to combining weak methods, or partially expanding universal algorithmic models of AI. The lack of “depth” of integration is evident from the fact that supporters of the above approaches prefer to oppose them to each other, criticizing the shortcomings of competitive approaches. Here we are talking more about the development of a new approach that takes into account the main previously obtained results and ideas at a much deeper conceptual level (although it is not always easy to establish a connection between different approaches).
It is necessary to start with the simplest models in the case of unlimited resources; make sure they are universal or establish what is missing to achieve it, which can be taken into account later. Next, you should consider universal models with a limit on computing resources. Such models can also be relatively simple, but should include self-optimization. Next, a priori information about the properties of the world (the most common of which will determine the characteristics of a cognitive architecture) should be introduced to reduce the time of the formation of AI, that is, to acquire autonomy.
(McCarthy, 2005) McCarthy J. The Future of AI — A Manifesto // AI Magazine. 2005. V. 26. No 4. P. 39.
(Nilsson, 2005a) Nilsson NJ Reconsiderations // AI Magazine. 2005. V. 26. No 4. P. 36–38.
(Nilsson, 2005b) Nilsson NJ Human-Level Artificial Intelligence? Be Serious! // AI Magazine. 2005. V. 26. No 4. P. 68–75.
(Brachman, 2005) Brachman R. Getting Back to “The Very Idea” // AI Magazine. 2005. V. 26. No 4. P. 48–50.
(Bobrow, 2005) Bobrow DG AAAI: It's Time for Large-Scale Systems // AI Magazine. 2005. V. 26. No 4. P. 40–41.
(Cassimatis et al., 2006) Cassimatis N., Mueller ET, Winston PH Achieving Human Intelligence through Integrated Systems and Research // AI Magazine. 2006. V. 27. No 2. P. 12-14.
(Langley, 2006) Langley P. Cognitive Architectures and General Intelligent Systems // AI Magazine. 2006. V. 27. No 2. P. 33–44.
(Cassimatis, 2006) Cassimatis NL A Cognitive Substrate for Achieving Human-Level Intelligence // AI Magazine. 2006. V. 27. No 2. P. 45–56.
(Jones and Wray, 2006) Jones RM, Wray RE Comparative Analysis of Frameworks for Knowledge-Intensive Intelligent Agents // AI Magazine. 2006. V. 27. No 2. P. 57-70.
(Cohen 2005) If What? // AI Magazine. 2005. V. 26. No 4. P. 61–67.
(Hall, 2008) J Storrs Hall. Engineering Utopia // Frontiers in Artificial Intelligence and Applications (Proc. 1st AGI Conference). 2008. V. 171. P. 460–467.
(Pankov, 2008) Pankov S. A computational approximation to the AIXI model // Frontiers in Artificial Intelligence and Applications (Proc. 1st AGI Conference). 2008. V. 171. P. 256–267.
(Duch et al., 2008) Duch W., Oentaryo RJ, Pasquier M. Cognitive Architectures: Where Do We Go From Here // Frontiers in Artificial Intelligence and Applications (Proc. 1st AGI Conference). 2008. V. 171. P. 122–136.
(Yudkowsky, 2011) Yudkowsky E. Complex Value Systems n Friendly AI // Proc. Artificial General Intelligence - 4th International Conference, AGI 2011, Mountain View, CA, USA, August 3-6, 2011. Lecture Notes in Computer Science 6830. Springer. 2011. p. 388–393.
(Kurzweil, 2005) Kurzweil R. The Singularity is Near. Viking, 2005.
(Solomonoff, 1964) Solomonoff RJ A formal theory of inductive inference: parts 1 and 2. Information and Control. 1964. V. 7. P. 1–22, 224–254.
(Schmidhuber, 2003) Schmidhuber J. The New AI: General & sound & relevant for physics. Technical Report TR IDSIA-04-03, Version 1.0, cs.AI/0302012 v1, IDSIA. 2003
(Hutter, 2001) Hutter M. Towards a universal theory of artificial intelligence based on algorithms and algorithms. In Proc. 12th European Conf. on Machine Learning (ECML-2001), volume 2167 of LNAI, Springer, Berlin. 2001.
(Hutter, 2005) Hutter M. Universal Artificial Intelligence. Sequential Decisions Based on Algorithmic Probability / Springer. 2005. 278 p.
(Wang, 2007) Wang P. The Logic of Intelligence // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. p. 31–62.
(Schmidhuber, 2007) Schmidhuber J. Gödel Machines: Fully Self-Referential Optimal Universal Self-improvers // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. P. 199–226.
(Hutter, 2007) Hutter M. Universal Algorithmic Intelligence: A Mathematical Top → Down Approach // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. p. 227–290.
(Goertzel and Pennachin, 2007) Goertzel B., Pennachin C. The Novamente Artificial Intelligence Engine // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. pp. 63–130.
(Garis, 2007) Hugo de Garis. Artificial Brains // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. p. 159–174
(Red'ko, 2007) Red'ko VG The Natural Way to Artificial Intelligence // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. P. 327–352.
A brief analysis of existing approaches to strong AI.
Cognitive architecture.
When creating a strong AI, it is natural to reproduce, if not all the details of the work of the human brain, then at least those functions that it performs. Otherwise, it is very difficult to be sure that it is the intellect that is created. It is this goal that cognitive architectures pursue, which combine functions such as learning, memory, planning, etc., that is, all (or almost all) of what is in natural intelligence. This makes cognitive architectures so attractive and popular.
However, in itself, the desire to endow the computer with all the same cognitive functions that a person has does not say how to do it correctly. As a result, a lot of cognitive architectures have been developed to date, a number of which are often positioned as a way to build a strong AI. These include, in particular, such popular architectures as “builders of strong AI”, such as Soar and ACT-R.
Many architectures often start from the phenomenology of higher cognitive functions of the human mind. However, due to the lack of a complete understanding of the nature of these functions and the requirements for them, their implementation is largely arbitrary.
Often, even the construction of such architectures is carried out within the framework of the traditional symbolic approach that models only the “tip of the iceberg” of human thinking. Nevertheless, an attempt is often made to construct architectures that reproduce not only high-level, but also low-level functions (so-called emergent architectures). Moreover, AI researchers are well aware of the need to combine symbolic and subsymbol levels and the development of hybrid architectures, as well as the need to build embodied systems (which are key, in particular, to obtain the semantic basis of concepts), which are highly problematic in purely symbolic architectures (see [ Duch et al., 2008] as a review).
Nevertheless, it is noted [Duch et al., 2008], which is very rarely possible to use for solving real problems, not to mention scaling to the level of autonomous behavior in the real world. So why did cognitive architectures not lead to significant progress in the field of strong AI? The answer to this question has already been given above.
These systems are most likely doomed to non-universality, since they are assembled from weak components. This apparently also applies to such systems that were initially positioned as general intelligence systems, such as Novamente (described in [Goertzel and Pennachin, 2007]). Of course, the possibility of introducing the property of universality as an extension of a particular architecture is not excluded (after all, the universality of intelligence can hardly be ascribed to most animals, which means it appeared as an evolutionary superstructure over more particular forms of intelligence). Nevertheless, this way seems to us more laborious and less optimal.
Approach based on resource constraints.
This approach is based on the following definition given by P. Wang [Wang, 2007]:
Whereas adaptation (as the ability to learn from experience) is a fairly common requirement, while the main features of the approach are derived from a lack of resources and knowledge (because when resources are and there is enough information, not quite intellectual methods can be used). As a result, within this approach, a categorical logic variant is built to take into account the fuzziness of knowledge, and a distributed knowledge manipulation system is proposed that takes into account the limited computational resources.
In this case, the author proposes to separate the concepts of "intellectual" and "effectively intellectual." Such a division seems to be quite fair and reflects the intuitive impression that, for example, a chess program operating by the “brute force” method is not intellectual in the same sense as the intellectual chess player.
Although it is possible to agree with the principle of effective intelligence, this particular approach can hardly become the basis for building FIS: it omits those aspects of intelligence that are revealed in universal algorithmic models and in cognitive architectures. In other words, the thesis about the need for resource constraints does not say how to enter them correctly.
In particular, this is evident from the fact that P. Wang introduced as a fundamental principle also the lack of knowledge of the agent. The lack of knowledge, of course, is important, but it is fully taken into account in (criticized by Wang) models of universal algorithmic intelligence, which include not only a search in the action space, but also a universal inductive conclusion, for which taking into account the fuzziness of knowledge is not a fundamental principle, but only heuristics to simplify sorting models (which will be demonstrated later).
As a result, within the framework of this approach, only a private cognitive architecture has been developed, which does not have any fundamental advantages over others, although systematically following the principles of limited resources also has considerable heuristic power.
')
Universal algorithmic intelligence.
The very idea of ​​this approach has been known for a long time, but it received recognition relatively recently mainly through the work of [Hutter, 2001], [Schmidhuber, 2003] and other works of these authors. Within its framework, the focus is on the Solomon’s model of universal induction, included in the system of choice of actions in the environment to maximize a certain estimating function.
Here, the analysis begins with a simple universal model, on which no resource constraints are imposed. The first step of our approach is similar, since we believe that the universality property is highly desirable to immediately introduce into the model of a universal AI and to maintain this property during the development of the model, which is carried out by introducing resource constraints.
In modern versions of the approaches under consideration, resource constraints are also introduced, but with preservation of the maximum impartiality of the universal AI, which allows us to construct general models of self-optimization.
This consideration of resource constraints, however, is not quite sufficient. It can be said that it requires the reproduction of the whole evolution, which also began as a universal self-optimizing search without any a priori information. It is obvious that the formation of such universal intelligence could be accomplished in the foreseeable time, it is necessary to lay in it as a sufficiently large amount of a priori information about the structure of the external world, and heuristics to reduce the search for models and actions. These heuristics can just be gleaned from the phenomenology of the cognitive functions of natural intelligence. On the other hand, it is not rational to manually build too much specific knowledge in a strong AI that he can learn on his own (which is what projects such as Cyc sin for). Obviously, it is necessary to achieve an optimal compromise between these two extremes.
In addition, a separate issue for discussion is whether or not the presented models are universal. For this, it is necessary to carefully compare the hypothetical capabilities of these models with the capabilities of a person. In part, such comparisons are made (for example, [Hutter, 2005]), although they cannot be called indisputable or exhaustive. Nevertheless, doubts about the true universality of these models can be put forward, which will be shown in the analysis of our own model of universal algorithmic intelligence.
Now we note only one of these doubts, which is that intelligence only in the zero approximation can be reduced to maximizing an a priori given objective function. After all, if, say, the task of intelligence is to ensure survival, then an a priori given objective function (based, say, on emotional evaluations) can only be a rough heuristic approximation of the goal of survival. This means the necessity of the existence of special mechanisms that allow in some way to refine the objective function in ontogenesis. Here you can make the following analogy with chess. Let a single intelligent agent play only one game. Having limited computing resources, he cannot perform a full search of options to predict victory or defeat. Being born with a minimum of a priori knowledge of the world, it cannot have a complex objective function that would effectively cut off unpromising options on the game tree. The initial objective function can rely only on some directly perceived stimuli, say the total force of the figures (giving a feeling of pain and pleasure when you lose your figure or eat an opponent's figure). In the process of growing up (a game), an agent can build more complex concepts, but on his own (without having lived life entirely), he cannot, in principle, determine how the objective function can be improved based on these concepts. This information, however, can be provided by other agents, but only if there is some good mechanism for modifying the objective function. This aspect is also related to the problem of friendly AI ...
Approach based on learning objective functions.
The problem of learning objective functions is sometimes considered as fundamental in building strong AI (or, more precisely, friendly AI [Yudkowsky, 2011]). Within the framework of this approach, it is rightly noted that maximizing an a priori objective function is insufficient for artificial intelligence to be universal, especially in terms of effective (and desired) interaction with the social environment, which is the same element of objective reality as the physical environment.
The problem of endowing AI with the ability to modify its own objective function is non-trivial due to the fact that it is not clear how the objective function can be optimized, if not under the control of another objective function (or some other a priori mechanisms). The importance of the possibility of modifying an objective function is connected not only with the fact that it is necessary for the full versatility of the agent, but also with the fact that the AI ​​seeking to maximize the a priori objective function may well find such actions optimal from the point of view of this function that will be extremely undesirable for people [Yudkowsky, 2011]. Although the importance of these aspects is indisputable, considering them outside of specific models of universal intelligence does not allow one to chart a way to create a strong AI (or rather, it sets some limitations on the way it is created), therefore this approach should be considered complementary to other approaches. The possibility of modifying the objective function must be provided in the architecture of the universal intelligent agent, although in general this can be considered at the same level as other cognitive functions, namely, as a specific heuristic of increasing the development efficiency of the "infant" AI to the level of "adult" AI.
Adaptive behavior, self-organization and bionics in general.
There is a large line of research in the field of strong AI, associated with the bionic approach. It highlights attempts (see, for example, [Garis, 2007] [Red'ko, 2007]) to model the brain at different levels of detail, to reproduce adaptive behavior, starting from its simplest forms to more complex ones, modeling evolution, self-organization as a whole. Often this approach is imitative in nature and rather rigidly opposed to the algorithmic approach, which is not deep enough. In particular, various simulation models of evolution and self-organization do not lead to unlimited development for the simple reason that their authors do not even try to consider issues related to the computational complexity of the optimization problems to be solved and the algorithmic completeness of those forms of behavior that can in principle be obtained during this modeling. Because of this, it is highly doubtful that the bionic approach itself can lead to the creation of a strong AI. However, at the same time, it can be an important source of productive ideas, which would be too wasteful to ignore.
Findings.
As you can see, different existing approaches to strong AI do not so much contradict each other, as they consider different aspects of the problem of universal AI, and therefore it is necessary to combine them. Naturally, there are many integration approaches that attempt to synthesize various existing systems and methods, so the idea of ​​integration is generally not new. However, this integration is often limited to combining weak methods, or partially expanding universal algorithmic models of AI. The lack of “depth” of integration is evident from the fact that supporters of the above approaches prefer to oppose them to each other, criticizing the shortcomings of competitive approaches. Here we are talking more about the development of a new approach that takes into account the main previously obtained results and ideas at a much deeper conceptual level (although it is not always easy to establish a connection between different approaches).
It is necessary to start with the simplest models in the case of unlimited resources; make sure they are universal or establish what is missing to achieve it, which can be taken into account later. Next, you should consider universal models with a limit on computing resources. Such models can also be relatively simple, but should include self-optimization. Next, a priori information about the properties of the world (the most common of which will determine the characteristics of a cognitive architecture) should be introduced to reduce the time of the formation of AI, that is, to acquire autonomy.
Literature.
(McCarthy, 2005) McCarthy J. The Future of AI — A Manifesto // AI Magazine. 2005. V. 26. No 4. P. 39.
(Nilsson, 2005a) Nilsson NJ Reconsiderations // AI Magazine. 2005. V. 26. No 4. P. 36–38.
(Nilsson, 2005b) Nilsson NJ Human-Level Artificial Intelligence? Be Serious! // AI Magazine. 2005. V. 26. No 4. P. 68–75.
(Brachman, 2005) Brachman R. Getting Back to “The Very Idea” // AI Magazine. 2005. V. 26. No 4. P. 48–50.
(Bobrow, 2005) Bobrow DG AAAI: It's Time for Large-Scale Systems // AI Magazine. 2005. V. 26. No 4. P. 40–41.
(Cassimatis et al., 2006) Cassimatis N., Mueller ET, Winston PH Achieving Human Intelligence through Integrated Systems and Research // AI Magazine. 2006. V. 27. No 2. P. 12-14.
(Langley, 2006) Langley P. Cognitive Architectures and General Intelligent Systems // AI Magazine. 2006. V. 27. No 2. P. 33–44.
(Cassimatis, 2006) Cassimatis NL A Cognitive Substrate for Achieving Human-Level Intelligence // AI Magazine. 2006. V. 27. No 2. P. 45–56.
(Jones and Wray, 2006) Jones RM, Wray RE Comparative Analysis of Frameworks for Knowledge-Intensive Intelligent Agents // AI Magazine. 2006. V. 27. No 2. P. 57-70.
(Cohen 2005) If What? // AI Magazine. 2005. V. 26. No 4. P. 61–67.
(Hall, 2008) J Storrs Hall. Engineering Utopia // Frontiers in Artificial Intelligence and Applications (Proc. 1st AGI Conference). 2008. V. 171. P. 460–467.
(Pankov, 2008) Pankov S. A computational approximation to the AIXI model // Frontiers in Artificial Intelligence and Applications (Proc. 1st AGI Conference). 2008. V. 171. P. 256–267.
(Duch et al., 2008) Duch W., Oentaryo RJ, Pasquier M. Cognitive Architectures: Where Do We Go From Here // Frontiers in Artificial Intelligence and Applications (Proc. 1st AGI Conference). 2008. V. 171. P. 122–136.
(Yudkowsky, 2011) Yudkowsky E. Complex Value Systems n Friendly AI // Proc. Artificial General Intelligence - 4th International Conference, AGI 2011, Mountain View, CA, USA, August 3-6, 2011. Lecture Notes in Computer Science 6830. Springer. 2011. p. 388–393.
(Kurzweil, 2005) Kurzweil R. The Singularity is Near. Viking, 2005.
(Solomonoff, 1964) Solomonoff RJ A formal theory of inductive inference: parts 1 and 2. Information and Control. 1964. V. 7. P. 1–22, 224–254.
(Schmidhuber, 2003) Schmidhuber J. The New AI: General & sound & relevant for physics. Technical Report TR IDSIA-04-03, Version 1.0, cs.AI/0302012 v1, IDSIA. 2003
(Hutter, 2001) Hutter M. Towards a universal theory of artificial intelligence based on algorithms and algorithms. In Proc. 12th European Conf. on Machine Learning (ECML-2001), volume 2167 of LNAI, Springer, Berlin. 2001.
(Hutter, 2005) Hutter M. Universal Artificial Intelligence. Sequential Decisions Based on Algorithmic Probability / Springer. 2005. 278 p.
(Wang, 2007) Wang P. The Logic of Intelligence // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. p. 31–62.
(Schmidhuber, 2007) Schmidhuber J. Gödel Machines: Fully Self-Referential Optimal Universal Self-improvers // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. P. 199–226.
(Hutter, 2007) Hutter M. Universal Algorithmic Intelligence: A Mathematical Top → Down Approach // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. p. 227–290.
(Goertzel and Pennachin, 2007) Goertzel B., Pennachin C. The Novamente Artificial Intelligence Engine // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. pp. 63–130.
(Garis, 2007) Hugo de Garis. Artificial Brains // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. p. 159–174
(Red'ko, 2007) Red'ko VG The Natural Way to Artificial Intelligence // in Artificial General Intelligence. Cognitive Technologies, B. Goertzel and C. Pennachin (Eds.). Springer. 2007. P. 327–352.
Source: https://habr.com/ru/post/145467/
All Articles