Search

The research field of AI is concerned with devising theories, methods, and workflows for producing software artifacts which behave as intelligent subjects. Evidently, intelligence, as the property of an agent, is not of necessity inherited from the methods used to construct it: that a car has been assembled by robots does not make it a robot. Unfortunately, even this obvious distinction can sometimes be erased in some prominent published work. To wit: the statement, “an agent that performs sufficiently well on a sufficiently wide range of tasks is classified as intelligent” was recently published by DeepMind [273] to give context to a paper claiming to have developed “the first deep RL agent that outperforms the standard human benchmark on all 57 Atari games” [14]. This invites the inference that the range of the tasks (57 games) that have been achieved warrants calling the advertised agent ‘intelligent’. However, careful reading of the paper reveals that the authors have in fact developed 57 different agents. Granted, this was achieved using the same development method and system architecture, but 57 agents were nonetheless trained, rather than the claimed single agent. Here is a prime example of distilled confusion: a property (applicability to 57 tasks) of one construction method (instantiating the Agent57 system architecture) has just been ‘magically’ transferred to some 57 artifacts produced by the method.

The failure of GOFAI created a vacancy for a new guiding philosophy for AI.

The system architecture presented in Chapter controls (i.e., sustains and constrains) the invocation of the inference methods introduced in Chapter. In this chapter, we describe the methods of higher level inference in more detail.

In this chapter, semantic closure meets system engineering: we describe how SCL systems can be constructed and controlled in practice, casting a developmental perspective on automation which we call ‘2nd order automation engineering’. Let us first give context to our objective, starting with a quote from Bundy and McNeil [40], who described in 2006 what they considered to be ‘a major goal of artificial intelligence research over the next 50 years’.

The rise of civilization is synonymous with the creation of tools that extend the intellectual and physical reach of human beings [133]. The pinnacle of such endeavours is to replicate the flexible reasoning capacity of human intelligence within a machine, making it capable of performing useful work on command, despite the complexity and adversity of the real world. In order to achieve such Artificial Intelligence (AI), a new approach is required: traditional symbolic AI has long been known to be too rigid to model complex and noisy phenomena and the sample-driven approach of Deep Learning cannot scale to the long-tailed distributions of the real world. In this book, we describe a new approach for building a situated system that reflects upon its own reasoning and is capable of making decisions in light of its limited knowledge and resources. This reflective reasoning process addresses the vital safety issues that inevitably accompany open-ended reasoning: the system must perform its mission within a specifiable operational envelope.

Machine learning excels at inducing mappings from data, but struggles to induce causal hierarchies. In contrast, symbolic reasoning (in particular, when considered as an expression language) can represent any form of domain knowledge and can index into code or data via pattern matching.

Deep learning (DL) has emerged as the dominant branch of machine learning, becoming the state of the art for machine intelligence in various domains. As discussed in the previous chapter, this has led some researchers to believe that deep learning could hypothetically scale to achieve general intelligence. However, there is increasing consensus (e.g. [57, 210, 230]) that the techniques do not scale as well as was anticipated to harder problems. In particular, deep learning methods find their strength in automatically synthesizing distributed quantitative features from data. These features are useful insofar as they enable mostly reliable classification and regression, and in some limited cases also few- or zero-shot transfer to related tasks. However, it is increasingly questionable whether deep learning methods are appropriate for autonomous roles in environments that are not strongly constrained. While there are still countless use-cases for narrow artificial intelligence, many of the truly transformative use-cases can only be realized by general intelligence

Let us recall that, from a pragmatic perspective, AI is nothing more or less than a tool for implementing a new leap in automation. This pragmatic perspective on AI is the one that matters in the current world economy, and therefore will necessarily receive primacy for development.

Reinforcement learning was historically established as a descriptive model of learning in animals [234], [324], [32], [279] then recast as a framework for optimal control [331].

In the domain of intelligent systems the management of system resources is typically called "attention". Attention mechanisms exist because even environments of moderate complexity are a source of vastly more information than available cognitive resources of any known intelligence can handle. Cognitive resource management has not been of much concern in artificial intelligence (AI) work that builds relatively simple systems for particular targeted problems. For systems capable of a wide range of actions in complex environments, explicit management of time and cognitive resources is not only useful, it is a necessity. We have designed a general attention mechanism for intelligent systems. While a full implementation remains to be realized, the architectural principles on which our work rests have already been implemented. Here we examine some prior work that we find relevant to engineered systems, describe our design, and how it derives from constructivist AI principles.

We present a cognitive architecture whose main constituents are allowed to grow through a situated experience in the world. Such an architectural growth is bootstrapped from a minimal initial knowledge and the architecture itself is built around the biologically-inspired notion of internal models. The key idea, supported by findings in cognitive neuroscience, is that the same internal models used in overt goal-directed action execution can be covertly re-enacted in simulation to provide a unifying explanation to a number of apparently unrelated individual and social phenomena, such as state estimation, action and intention understanding, imitation learning and mindreading. Thus, rather than reasoning over abstract symbols, we rely on the biologically plausible processes firmly grounded in the actual sensorimotor experience of the agent. The article describes how such internal models are learned in the first place, either through individual experience or by observing and imitating other skilled agents, and how they are used in action planning and execution. Furthermore, we explain how the architecture continuously adapts its internal agency and how increasingly complex cognitive phenomena, such as continuous learning, prediction and anticipation, result from an interplay of simpler principles. We describe an early evaluation of our approach in a classical AI problem-solving domain: the Sokoban puzzle.

Intuitively speaking, “self-programming” means the ability for a computer system to program its own actions. This notion is clearly related to Artificial Intelligence, and has been used by many researchers. Like many other high-level concepts, however, scrutiny shows that the term can be interpreted in several different ways. To make the discussion concrete and meaningful we introduce here a working definition of self-programming. In this definition we increase its concreteness while trying to keep the intuitive meaning of the concept. The activities of a computer system usually are considered to consist of atomic actions (which can also be called instructions, operations, behavior, or something else in different contexts). At any given moment the system’s primitive actions are in a finite and constant set A, meaning that they are distinct from each other, and can be enumerated. An action may take some input arguments, and produce some output arguments. The system can execute each of its actions

Dependable cyber-physical systems strive to deliver anticipative, multi-objective performance anytime, facing deluges of inputs with varying and limited resources. This is even more challenging for life-long learning rational agents as they also have to contend with the varying and growing know-how accumulated from experience. These issues are of crucial practical value, yet have been only marginally and unsatisfactorily addressed in AGI research. We present a value-driven computational model of anytime bounded rationality robust to variations of both resources and knowledge. It leverages continually learned knowledge to anticipate, revise and maintain concurrent courses of action spanning over arbitrary time scales for execution anytime necessary.

Four principal features of autonomous control systems are left both unaddressed and unaddressable by present-day engineering methodologies: (1) The ability to operate effectively in environments that are only partially known at design time; (2) A level of generality that allows a system to re-assess and re-define the fulfillment of its mission in light of unexpected constraints or other unforeseen changes in the environment; (3) The ability to operate effectively in environments of significant complexity; and (4) The ability to degrade gracefully—how it can continue striving to achieve its main goals when resources become scarce, or in light of other expected or unexpected constraining factors that impede its progress. We describe new methodological and engineering principles for addressing these shortcomings, that we have used to design a machine that becomes increasingly better at behaving in underspecified circumstances, in a goal-directed way, on the job, by modeling itself and its environment as experience accumulates. The work provides an architectural blueprint for constructing systems with high levels of operational autonomy in underspecified circumstances, starting from only a small amount of designer-specified code—a seed. Using value-driven dynamic priority scheduling to control the parallel execution of a vast number of lines of reasoning, the system accumulates increasingly useful models of its experience, resulting in recursive self-improvement that can be autonomously sustained after the machine leaves the lab, within the boundaries imposed by its designers. A prototype system named AERA has been implemented and demonstrated to learn a complex real-world task—real-time multimodal dialogue with humans—by on-line observation. Our work presents solutions to several challenges that must be solved for achieving artificial general intelligence.

In this paper we consider the issue of endowing an AGI system with decision-making capabilities for operation in real-world environments or those of comparable complexity. While action-selection is a critical function of any AGI system operating in the real-world, very few applicable theories or methodologies exist to support such functionality, when all necessary factors are taken into account. Decision theory and standard search techniques require several debilitating simplifications, including determinism, discrete state spaces, exhaustive evaluation of all possible future actions and a coarse grained representation of time. Due to the stochastic and continuous nature of real-world environments and inherent time-constraints, direct application of decision-making methodologies from traditional decision theory and search is not a viable option. We present predictive heuristics as a way to bridge the gap between the simplifications of decision theory and search, and the complexity of real-world environments.

Resource-boundedness must be carefully considered when designing and implementing artificial general intelligence (AGI) algorithms and architectures that have to deal with the real world. But not only must resources be represented explicitly throughout its design, an agent must also take into account their usage and the associated costs during reasoning and acting. Moreover, the agent must be intrinsically motivated to become progressively better at utilizing resources. This drive then naturally leads to effectiveness, efficiency, and curiosity. We propose a practical operational framework that explicitly takes into account resource constraints: activities are organized to maximally utilize an agent's bounded resources as well as the availability of a teacher, and to drive the agent to become progressively better at utilizing its resources. We show how an existing AGI architecture called AERA can function inside this framework. In short, the capability of AERA to perform self-compilation can be used to motivate the system to not only accumulate knowledge and skills faster, but also to achieve goals using less resources, becoming progressively more effective and efficient.

Many existing AGI architectures are based on the assumption of infinite computa- tional resources, as researchers ignore the fact that real-world tasks have time limits, and managing these is a key part of the role of intelligence. In the domain of intelligent systems the management of system resources is typically called “attention”. Attention mechanisms are necessary because all moderately complex environments are likely to be the source of vastly more information than could be processed in realtime by an intelligence’s available cognitive resources. Even if sufficient resources were available, attention could help make better use of them. We argue that attentional mechanisms are not only nice to have, for AGI architectures they are an absolute necessity. We examine ideas and concepts from cognitive psychology for creating intelligent resource management mechanisms and how these can be applied to engineered systems. We present a design for a general attention mechanism intended for imple- mentation in AGI architectures.

A key goal in designing an artificial intelligence capable of performing complex tasks is a mechanism that allows it to efficiently choose appropriate and relevant actions in a variety of situations and contexts. Nowhere is this more obvious than in the case of building a general intelligence, where the contextual choice and application of actions must be done in the presence of large numbers of alternatives, both subtly and obviously distinct from each other. We present a framework for action selection based on the concurrent activity of multiple forward and inverse models. A key characteristic of the proposed system is the use of simulation to choose an action: the system continuously simulates the external states of the world (proximal and distal) by internally emulating the activity of its sensors, adopting the same decision process as if it were actually operating in the world, and basing subsequent choice of action on the outcome of such simulations. The work is part of our larger effort to create new observation-based machine learning techniques. We describe our approach, an early implementation, and an evaluation in a classical AI problem-solving domain: the Sokoban puzzle.

We present an architectural approach to learning problem solving skills from demonstration, using internal models to represent problem-solving operational knowledge. Internal forward and inverse models are initially learned through active interaction with the environment, and then enhanced and finessed by observing expert teachers. While a single internal model is capable of solving a single goal-oriented task, it is their sequence that enables the system to hierarchically solve more complex task. Activation of models is goal-driven, and internal ”mental” simulations are used to predict and anticipate future rewards and perils and to make decisions accordingly. In this approach intelligent system behavior emerges as a coordinated activity of internal models over time governed by sound architectural principles. In this paper we report preliminary results using the game of Sokoban, where the aim is to learn goal-oriented patterns of model activations capable of solving the problem in various contexts.

The general intelligence of any autonomous system must in large part be measured by its ability to automatically learn new skills and integrate these with prior skills. Cognitive architectures addressing these topics are few and far between ‐ possibly because of their difficulty. We argue that architectures capable of diverse skill acquisition and integration, and real-time management of these, require an approach of modularization that goes well beyond the current practices, leading to a class of architectures we refer to as peewee-granule systems. The building blocks (modules) in such systems have simple operational semantics and result in architectures that are heterogeneous at the cognitive level but homogeneous at the computational level.

Certain necessary features of general intelligence are more system-wide than others; features such as attention, learning and temporal grounding are transversal in that they seem to affect a significant subset of all mental operation. We argue that such transversal features unavoidably impose fundamental constraints on the kinds of architectures and methodologies required for building artificially intelligent systems. Current component-based software practices fall short for building systems with transversal features: Artificial general intelligence efforts call for new system architectures and new methodologies, where transversal features must be taken into account from the very outset.

Giving synthetic agents human-like realtime turntaking skills is a challenging task. Attempts have been made to manually construct such skills, with systematic categorization of silences, prosody and other candidate turn-giving signals, and to use analysis of corpora to produce static decision trees for this purpose. However, for general-purpose turntaking skills which vary between individuals and cultures, a system that can learn them on-the-job would be best. We are exploring ways to use machine learning to have an agent learn proper turntaking during interaction. We have implemented a talking agent that continuously adjusts its turntaking behavior to its interlocutors based on realtime analysis of the other party's prosody. Initial results from experiments on collaborative, content-free dialogue show that, for a given subset of turn-taking conditions, our modular reinforcement learning techniques allow the system to learn to take turns in an efficient, human-like manner.