Ebook: Artificial General Intelligence 2008

The field of Artificial Intelligence (AI) was initially directly aimed at the construction of “thinking machines” – that is, computer systems with human-like general intelligence. But this task proved more difficult than expected. As the years passed, AI researchers gradually shifted focus to producing AI systems that intelligently approached specific tasks in relatively narrow domains.

In recent years, however, more and more AI researchers have recognized the necessity – and the feasibility – of returning to the original goal of the field. Increasingly, there is a call to focus less on highly specialized “narrow AI” problem solving systems, and more on confronting the difficult issues involved in creating “human-level intelligence,” and ultimately general intelligence that goes beyond the human level in various ways. “Artificial General Intelligence (AGI)”, as this renewed focus has come to be called, attempts to study and reproduce intelligence as a whole in a domain-independent way.

Encouraged by the recent success of several smaller-scale AGI-related meetings and special tracks at conferences, we took the initiative to organize the very first international conference on AGI. Our goal in doing so was to give researchers in the field an opportunity to present relevant research results and to exchange ideas on topics of common interest.

The response to AGI-08 was even stronger than we expected. We received many interesting papers addressing a wide range of AGI-relevant topics and exploring various kinds of theoretical and technical ideas. Given the complexity and difficulty of the problems the field is facing, we believe it is crucial to encourage the exchange of ideas and opinions with various degrees of maturity, ranging from position papers to descriptions of mature mathematical and cognitive AGI theories, and practical work with implemented AGI architectures.

In this collection, the AGI-08 papers are organized into three groups. The conference papers are divided into full-length papers (12 pages, with a few exceptions) and short position statements (5 pages). Also included are the papers presented in the post-conference workshop on the sociocultural, ethical and futurological implications of AGI.

We believe meetings like AGI-08 are important, not only because of their presentations and discussions, but also because of the potential they have to help catalyze the self-organization of a vital AGI research community. Together, we will continue to directly challenge one of the most difficult and essential problems in human history, the creation of human-level artificial general intelligence.

We thank the following referees for devoting their valuable time to reviewing the submitted papers: Sam Adams, James Anderson, Mike Anderson, Eric Baum, Mark Bickhard, Henry Brighton, Nick Cassimatis, Hernan Castro, L. Andrew Coward, Hugo de Garis, Wlodzislaw Duch, Richard Duro, David Friedlander, J. Storrs Hall, David Hart, Marcus Hutter, Cliff Joslyn, Randal Koene, Moshe Looks, Bruce MacLennan, Don Perlis, Matthias Scheutz, Juergen Schmidhuber, Lokendra Shastri, Boris Velichkovsky, Karin Verspoor, Paul Vogt, and Mark Waser.

We would also like to thank the other members of the AGI-08 Organizing Committee, Sidney D'Mello, Bruce Klein, and Lee McCauley, for the thought and effort they expended in preparing the conference; and Bruce Klein and Natasha Vita-More for their help with the post-conference workshop which resulted in a number of the papers contributed to this volume.

We consider the Distributed Cognition paradigm as a framework for implementing artificial components of human cognition. We take email/internet search as a setting of distributed cognition and describe an algorithm that intervenes and enhances a component of this distributed process. Results are presented demonstrating the effectiveness of the algorithm in interaction with an automaton simulating human search activity. We define a notion of synchronization with a non-random, non-Turing computation and argue that the given algorithm exhibits such synchronization behavior. Our results suggest that the framework may be useful for studying a class of non-Turing computation that is central to General Intelligence.

We present a mathematical model of interacting neuron-like units that we call Input Feedback Networks (IFN). Our model is motivated by a new approach to biological neural networks, which contrasts with current approaches (e.g. Layered Neural Networks, Perceptron etc.). Classification reasoning in IFN are accomplished by an iterative algorithm, and learning changes only structure. Feature relevance is determined during classification. Thus it emphasizes network structure over edge weights. IFNs are more flexible than previous approaches. In particular, integration of a new node can affect the outcome of existing nodes without modifying their prior structure. IFN can produce informative responses to partial inputs or when the networks are extended to other tasks. It also enables recognition of complex entities (e.g. images) from parts. This new model is promising for future contributions to integrated human-level intelligent applications due to its flexibility, dynamics and structural similarity to natural neuronal networks.

We introduce a general framework for reasoning with prioritized data by aggregation of distance functions, study some basic properties of the entailment relations that are obtained, and relate them to other approaches for maintaining uncertain information.

Representing uncertainty and reasoning with dynamically evolving systems are two related issues that are in the heart of many information systems. In this paper we show that these tasks can be successfully dealt with by incorporating distance semantics and non-deterministic matrices. The outcome is a general framework for capturing the principle of minimal change and providing a non-deterministic view of the domain of discourse. We investigate some properties of the entailment relations that are induced by this framework and demonstrate their usability in some test-cases.

This paper, the second in a series, provides the theory and formalisms for the implementation of an ethical control and reasoning system potentially suitable for constraining lethal actions in an autonomous robotic system. so that they fall within the bounds prescribed by the Laws of War and Rules of Engagement. It is based upon extensions to existing deliberative/reactive autonomous robotic architectures.

Understanding why the original project of Artificial Intelligence is widely regarded as a failure and has been abandoned even by most of contemporary AI research itself may prove crucial to achieving synthetic intelligence. Here, we take a brief look at some principles that we might consider to be lessons from the past five decades of AI. The author's own AI architecture – MicroPsi – attempts to contribute to that discussion.

Human brain is exceptionally complex and simple at the same time. Its extremely composite biological structure results itself in human everyday behavior that many people might consider rather simple than complex. In our research we will concentrate on the ways how a human brain can processes English and other human natural languages because taken in general sense the ability to speak English or other human languages is only serious distinguishing feature that rises humans over the rest of the world making a human an intellectual being. On the purpose of our research we consider natural language as naturally formed symbolic system completely independent of these symbols' physical nature that is a little more general than a common natural language definition. The principles of natural language processing in human brain are most important for us if we want to build equally powerful artificial general intelligence. We start with the features of human brain neurons and neural networks, and step by step create a computer model of human brain networks that is able to process and generate a reasonable speech. We can't give a detailed explanation of human brain functionality in this short article. Moreover, it is not our goal, and such research is not complete yet. The main result of our research is revealing the principles how tiny single neurons working together can produce intellectual-like behaviour that exhibits itself in proper speech comprehension and generation in accordance with current context.

With respect to genuine cognitive faculties, present synthetic characters inhabiting online virtual worlds are, to say the least, completely impaired. Current methods aimed at the creation of “immersive” virtual worlds only avatars and NPCs the illusion of mentality and, as such, will ultimately fail. Like behaviorism, this doomed approach focuses only on the inputs and outputs of virtual characters and ignores the rich mental structures that are essential for any truly realistic social environment. While this “deceptive” tactic may be suitable so long as a human is in the driver's seat compensating for the mental deficit, truly convincing autonomous synthetic characters must possess genuine mental states, which can only result from a formal theory of mind. We report here on our attempt to invent part of such a theory, one that will enable artificial agents to have and reason about the beliefs of others, resulting in characters that can predict and manipulate the behavior of even human players. Furthermore, we present the “embodiment” of our recent successes: Eddie, a four year old child in Second Life who can reason about his own beliefs to draw conclusions in a manner that matches human children his age.

Our goal is to understand human language use and create systems that can use human language fluently. We argue that to a achieve this goal, we must formulate all of the problems for language use from morphology up to pragmatics using the same cognitive substrate of reasoning and representation abilities. We propose such a substrate and described systems based on it. Our arguments, results with real-world systems and ongoing work suggest that the cognitive substrate enables a significant advance in the power of cognitive models and intelligent systems to use human language.

Prof. Hugo de Garis has recently received a 3 million RMB, 4 year grant to build China's first artificial brain, starting in 2008, that will consist of approximately 15,000 interconnected neural net modules, evolved one at a time in a special accelerator board [1] (which is 50 times faster than using an ordinary PC) to control the hundreds of behaviors of an autonomous robot. The approach taken in building this artificial brain is fast and cheap (e.g. $1500 for the FPGA board, $1000 for the robot, and $500 for the PC, a total of $3000), so we hope that other brain building groups around the world will copy this evolutionary engineering approach.

Cognitive architectures play a vital role in providing blueprints for building future intelligent systems supporting a broad range of capabilities similar to those of humans. How useful are existing architectures for creating artificial general intelligence? A critical survey of the state of the art in cognitive architectures is presented providing a useful insight into the possible frameworks for general intelligence. Grand challenges and an outline of the most promising future directions are described.

Every agent aspiring to human level intelligence, every AGI agent, must be capable of a theory of mind. That is, it must be able to attribute mental states, including intentions, to other agents, and must use such attributions in its action selection process. The LIDA conceptual and computational model of cognition offers an explanation of how theory of mind is accomplished in humans and some other animals, and suggests how this explanation could be implemented computationally. Here we describe how the LIDA version of theory of mind is accomplished, and illustrate it with an example taken from an experiment with monkeys, chosen for simplicity.

A series of hypotheses is proposed, connecting neural structures and dynamics with the formal structures and processes of probabilistic logic. First, a hypothetical connection is proposed between Hebbian learning in the brain and the operation of probabilistic term logic deduction. It is argued that neural assemblies could serve the role of logical terms; synapse-bundles joining neural assemblies could serve the role of first-order term-logic statements; and in this framework, Hebbian learning at the synaptic level would be expected to have the implicit consequence of probabilistic deduction at the logical statement level. A conceptual problem arises with this idea, pertaining to the brain's lack of a mechanism of “inference trails” as used to avoid circular reasoning in AI term logic inference systems; but it is explained how this problem may be circumvented if one posits an appropriate inference control mechanism. Finally, a brief discussion is given regarding the potential extension of this approach to handle more complex logical expressions involving variables – via the hypothesis of special neural structures mapping neural weights into neural inputs, hence implementing “higher order functions.”

A teaching methodology called Imitative-Reinforcement-Corrective (IRC) learning is described, and proposed as a general approach for teaching embodied non-linguistic AGI systems. IRC may be used with a variety of different learning algorithms, but it is particularly easily described in EC lingo. In these terms, it is a framework for automatically learning a procedure that generates a desired type of behavior, in which: a set of exemplars of the target behavior-type are utilized for fitness estimation; reinforcement signals from a human teacher are used for fitness evaluation; and the execution of candidate procedures may be modified by the teacher via corrections delivered in real-time. An example application of IRC to teach behaviors to AI-controlled artificial animals embodied in the Second Life virtual world is described in detail, including a review of the overall virtual-animal-control software architecture and how the integrative teaching/learning methodology fits into it. In this example application architecture, the learning algorithm may be toggled between hillclimbing and probabilistic evolutionary learning. Envisioned future applications are also discussed, including an application to embodied language learning applicable to agents in Second Life and other virtual worlds.

Learning theory and programs to date are inductively bounded: they can be described as “wind-up toys” which can only learn the kinds of things that their designers envisioned. We conjecture [1] that general intelligence involves an unbounded learning ability. VARIAC is an experimental cognitive architecture designed to learn by modifying and extending itself, including its ability to learn, so that it can learn to become a better learner.

Indefinite probabilities are a novel technique for quantifying uncertainty, which were created as part of the PLN (Probabilistic Logic Networks) logical inference engine, which is a key component of the Novamente Cognition Engine (NCE), an integrative AGI system. Previous papers have discussed the use of indefinite probabilities in the context of a variety of logical inference rules, but have omitted discussion of quantification. Here this gap is filled, and a mathematical procedure is provided allowing the propagation of indefinite probabilities through universal and existential quantifiers, and also through a variety of fuzzy quantifiers corresponding to natural language quantifiers (such as “few”, “many”, “a lot”, “hardly any”, etc.). Proper probabilistic handling of various quantifier transformation rules is also discussed. Together with the ideas in prior publications, and a forthcoming sequel paper on indefinite probabilities for intensional inference, these results allow probabilistic logic based on indefinite probabilities to be utilized for the full scope of inferences involved in intelligent reasoning. To illustrate the ideas and algorithms involved, we give two concrete examples: Halpern's “crooked lottery” predicate, and a commonsense syllogism that uses fuzzy quantifiers together with the standard PLN term logic deduction rule.

Rich computer simulations or quantitative models can enable an agent to realistically predict real-world behavior with precision and performance that is difficult to emulate in logical formalisms. Unfortunately, such simulations lack the deductive flexibility of techniques such as formal logics and so do not find natural application in the deductive machinery of commonsense or general purpose reasoning systems. This dilemma can, however, be resolved via a hybrid architecture that combines tableaux-based reasoning with a framework for generic simulation based on the concept of ‘molecular’ models. This combination exploits the complementary strengths of logic and simulation, allowing an agent to build and reason with automatically constructed simulations in a problem-sensitive manner.

Whereas symbol-based systems, like deductive reasoning devices, knowledge bases, planning systems, or tools for solving constraint satisfaction problems, presuppose (more or less) the consistency of data and the consistency of results of internal computations, this is far from being plausible in real-world applications, in particular, if we take natural agents into account. Furthermore in complex cognitive systems, that often contain a large number of different modules, inconsistencies can jeopardize the integrity of the whole system. This paper addresses the problem of resolving inconsistencies in hybrid cognitively inspired systems on both levels, in single processing modules and in the overall system. We propose the hybrid architecture I-Cog as a flexible tool, that is explicitly designed to reorganize knowledge constantly and use occurring inconsistencies as a non-classical learning mechanism.

One approach in pursuit of general intelligent agents has been to concentrate on the underlying cognitive architecture, of which Soar is a prime example. In the past, Soar has relied on a minimal number of architectural modules together with purely symbolic representations of knowledge. This paper presents the cognitive architecture approach to general intelligence and the traditional, symbolic Soar architecture. This is followed by major additions to Soar: non-symbolic representations, new learning mechanisms, and long-term memories.

Inhabiting the complex and dynamic environments of modern computer games with autonomous agents capable of intelligent timely behaviour is a significant research challenge. We illustrate this using our own attempts to build a practical agent architecture on a logicist foundation. In the ANDI-Land adventure game concept players solve puzzles by eliciting information from computer characters through natural language question answering. While numerous challenges immediately presented themselves, they took on a form of concrete and accessible problems to solve, and we present some of our initial solutions. We conclude that games, due to their demand for human-like computer characters with robust and independent operation in large simulated worlds, might serve as excellent test beds for research towards artificial general intelligence.

An artificial system that achieves human-level performance on open-domain tasks must have a huge amount of knowledge about the world. We argue that the most feasible way to construct such a system is to let it learn from the large collections of text, images, and video that are available online. More specifically, the system should use a Bayesian probability model to construct hypotheses about both specific objects and events, and general patterns that explain the observed data.

Universal induction solves in principle the problem of choosing a prior to achieve optimal inductive inference. The AIXI theory, which combines control theory and universal induction, solves in principle the problem of optimal behavior of an intelligent agent. A practically most important and very challenging problem is to find a computationally efficient (if not optimal) approximation for the optimal but incomputable AIXI theory. We propose such an approximation based on a Monte Carlo algorithm that samples programs according to their algorithmic probability. The approach is specifically designed to deal with real world problems (the agent processes observed data and makes plans over range of divergent time scales) under limited computational resources.

We present a set of cognitive phenomena that should be exhibited by a generally intelligent system. To date, we know of few systems that address more than a handful of these phenomena, and none that are able to explain all of them. Thus, these phenomena should motivate a system's design, test its generality, and can be used to point out fundamental shortcomings. The phenomena encourage autonomous learning, development of representations, and domain independence, which we argue are critical for general intelligence.