Enhancing the performance of an agent-based manufacturing system through learning and forecasting

Agent-based technology has been identified as an important approach for developing next generation manufacturing systems. One of the key techniques needed for implementing such advanced systems will be learning. This paper first discusses learning issues in agent-based manufacturing systems and reviews related approaches, then describes how to enhance the performance of an agent-based manufacturing system through learning from history (based on distributed case-based learning and reasoning) and learning from the future (through system forecasting simulation). Learning from history is used to enhance coordination capabilities by minimizing communication and processing overheads. Learning from the future is used to adjust promissory schedules through forecasting simulation, by taking into account the shop floor interactions, production and transportation time. Detailed learning and reasoning mechanisms are described and partial experimental results are presented.     

The Power of Self-Directed Learning

This article studies self-directed learning, a variant of the on-line (or incremental) learning model in which the learner selects the presentation order for the instances. Alternatively, one can view this model as a variation of learning with membership queries in which the learner is only charged for membership queries for which it could not predict the outcome. We give tight bounds on the complexity of self-directed learning for the concept classes of monomials, monotone DNF formulas, and axis-parallel rectangles in {0, 1, , n – 1} ^d . These results demonstrate that the number of mistakes under self-directed learning can be surprisingly small. We then show that learning complexity in the model of self-directed learning is less than that of all other commonly studied on-line and query learning models. Next we explore the relationship between the complexity of self-directed learning and the Vapnik-Chervonenkis (VC-)dimension. We show that, in general, the VC-dimension and the self-directed learning complexity are incomparable. However, for some special cases, we show that the VC-dimension gives a lower bound for the self-directed learning complexity. Finally, we explore a relationship between Mitchell's version space algorithm and the existence of self-directed learning algorithms that make few mistakes.     

Prequential and Cross-Validated Regression Estimation

Prequential model selection and delete-one cross-validation are data-driven methodologies for choosing between rival models on the basis of their predictive abilities. For a given set of observations, the predictive ability of a model is measured by the model's accumulated prediction error and by the model's average-out-of-sample prediction error, respectively, for prequential model selection and for cross-validation. In this paper, given i.i.d. observations, we propose nonparametric regression estimators—based on neural networks—that select the number of hidden units (or neurons) using either prequential model selection or delete-one cross-validation. As our main contributions: (i) we establish rates of convergence for the integrated mean-squared errors in estimating the regression function using off-line or batch versions of the proposed estimators and (ii) we establish rates of convergence for the time-averaged expected prediction errors in using on-line versions of the proposed estimators. We also present computer simulations (i) empirically validating the proposed estimators and (ii) empirically comparing the proposed estimators with certain novel prequential and cross-validated mixture regression estimators.     

On the Complexity of Function Learning

The majority of results in computational learning theory are concerned with concept learning, i.e. with the special case of function learning for classes of functions with range {0, 1}. Much less is known about the theory of learning functions with a larger range such as or . In particular relatively few results exist about the general structure of common models for function learning, and there are only very few nontrivial function classes for which positive learning results have been exhibited in any of these models.We introduce in this paper the notion of a binary branching adversary tree for function learning, which allows us to give a somewhat surprising equivalent characterization of the optimal learning cost for learning a class of real-valued functions (in terms of a max-min definition which does not involve any learning model).Another general structural result of this paper relates the cost for learning a union of function classes to the learning costs for the individual function classes.Furthermore, we exhibit an efficient learning algorithm for learning convex piecewise linear functions from ^d into . Previously, the class of linear functions from ^d into was the only class of functions with multidimensional domain that was known to be learnable within the rigorous framework of a formal model for online learning.Finally we give a sufficient condition for an arbitrary class of functions from into that allows us to learn the class of all functions that can be written as the pointwise maximum ofk functions from . This allows us to exhibit a number of further nontrivial classes of functions from into for which there exist efficient learning algorithms.     

Separating complexity classes related to bounded alternating ω-branching programs

We develop a theory of communication within branching programs that provides exponential lower bounds on the size of branching programs that are bounded alternating. Our theory is based on the algebraic concept of -branching programs, : , a semiring homomorphism, that generalizes ordinary branching programs, -branching programs [M2] andMOD _p-branching programs [DKMW].Due to certain exponential lower and polynomial upper bounds on the size of bounded alternating -branching programs we are able to separate the corresponding complexity classesN _ba ,co-N _{ba ba} , andMOD _p - _ba ,p prime, from each other, and from that classes corresponding to oblivious linear length-bounded branching programs investigated in the past.     

Selective Sampling Using the Query by Committee Algorithm

We analyze the query by committee algorithm, a method for filtering informative queries from a random stream of inputs. We show that if the two-member committee algorithm achieves information gain with positive lower bound, then the prediction error decreases exponentially with the number of queries. We show that, in particular, this exponential decrease holds for query learning of perceptrons.     

Rapid Concept Learning for Mobile Robots

Concept learning in robotics is an extremely challenging problem: sensory data is often high dimensional, and noisy due to specularities and other irregularities. In this paper, we investigate two general strategies to speed up learning, based on spatial decomposition of the sensory representation, and simultaneous learning of multiple classes using a shared structure. We study two concept learning scenarios: a hallway navigation problem, where the robot has to induce features such as opening or wall. The second task is recycling, where the robot has to learn to recognize objects, such as a trash can. We use a common underlying function approximator in both studies in the form of a feedforward neural network, with several hundred input units and multiple output units. Despite the high degree of freedom afforded by such an approximator, we show the two strategies provide sufficient bias to achieve rapid learning. We provide detailed experimental studies on an actual mobile robot called PAVLOV to illustrate the effectiveness of this approach.     

From Conditional Events to Conditional Measures: A New Axiomatic Approach

Our starting point is a definition of conditional event EH which differs from many seemingly similar ones adopted in the relevant literature since 1935, starting with de Finetti. In fact, if we do not assign the same third value u (undetermined) to all conditional events, but make it depend on EH, it turns out that this function t(EH) can be taken as a general conditional uncertainty measure, and we get (through a suitable – in a sense, compulsory – choice of the relevant operations among conditional events) the natural axioms for many different (besides probability) conditional measures.     

Rapid Concept Learning for Mobile Robots

Concept learning in robotics is an extremely challenging problem: sensory data is often high-dimensional, and noisy due to specularities and other irregularities. In this paper, we investigate two general strategies to speed up learning, based on spatial decomposition of the sensory representation, and simultaneous learning of multiple classes using a shared structure. We study two concept learning scenarios: a hallway navigation problem, where the robot has to induce features such as opening or wall. The second task is recycling, where the robot has to learn to recognize objects, such as a trash can. We use a common underlying function approximator in both studies in the form of a feedforward neural network, with several hundred input units and multiple output units. Despite the high degree of freedom afforded by such an approximator, we show the two strategies provide sufficient bias to achieve rapid learning. We provide detailed experimental studies on an actual mobile robot called PAVLOV to illustrate the effectiveness of this approach.     

Parallel complexity of logical query programs

We consider the parallel time complexity of logic programs without function symbols, called logical query programs, or Datalog programs. We give a PRAM algorithm for computing the minimum model of a logical query program, and show that for programs with the polynomial fringe property, this algorithm runs in time that is logarithmic in the input size, assuming that concurrent writes are allowed if they are consistent. As a result, the linear and piecewise linear classes of logic programs are inN C. Then we examine several nonlinear classes in which the program has a single recursive rule that is an elementary chain. We show that certain nonlinear programs are related to GSM mappings of a balanced parentheses language, and that this relationship implies the polynomial fringe property; hence such programs are inN C Finally, we describe an approach for demonstrating that certain logical query programs are log space complete forP, and apply it to both elementary single rule programs and nonelementary programs.Supported by NSF Grant IST-84-12791, a grant of IBM Corporation, and ONR contract N00014-85-C-0731.     

Complexity and computability of solutions to linear programming systems

Through key examples and constructs, exact and approximate, complexity, computability, and solution of linear programming systems are reexamined in the light of Khachian's new notion of (approximate) solution. Algorithms, basic theorems, and alternate representations are reviewed. It is shown that the Klee-Minty example hasnever been exponential for (exact) adjacent extreme point algorithms and that the Balinski-Gomory (exact) algorithm continues to be polynomial in cases where (approximate) ellipsoidal centered-cutoff algorithms (Levin, Shor, Khachian, Gacs-Lovasz) are exponential. By model approximation, both the Klee-Minty and the new J. Clausen examples are shown to be trivial (explicitly solvable) interval programming problems. A new notion of computable (approximate) solution is proposed together with ana priori regularization for linear programming systems. New polyhedral constraint contraction algorithms are proposed for approximate solution and the relevance of interval programming for good starts or exact solution is brought forth. It is concluded from all this that the imposed problem ignorance of past complexity research is deleterious to research progress on computability or efficiency of computation.This research was partly supported by Project NR047-071, ONR Contract N00014-80-C-0242, and Project NR047-021, ONR Contract N00014-75-C-0569, with the Center for Cybernetic Studies, The University of Texas at Austin.     

The dynamic clusters method in nonhierarchical clustering

Given a finite setE R ⁿ, the problem is to find clusters (or subsets of similar points inE) and at the same time to find the most typical elements of this set. An original mathematical formulation is given to the problem. The proposed algorithm operates on groups of points, called samplings (samplings may be called multiple centers or cores); these samplings adapt and evolve into interesting clusters. Compared with other clustering algorithms, this algorithm requires less machine time and storage. We provide some propositions about nonprobabilistic convergence and a sufficient condition which ensures the decrease of the criterion. Some computational experiments are presented.     

Interval routing schemes

In this paper the problem of routing messages along shortest paths in a distributed network without using complete routing tables is considered. In particular, the complexity of deriving minimum (in terms of number of intervals) interval routing schemes is analyzed under different requirements. For all the cases considered NP-hardness proofs are given, while some approximability results are provided. Moreover, relations among the different cases considered are studied.This work was supported by the EEC ESPRIT II Basic Research Action Program under Contract No. 7141 Algorithms and Complexity II, by the EEC Human Capital and Mobility MAP project, and by the Italian MURST 40% project Algoritmi, Modelli di Calcolo e Strutture Informative.     

Reasoning about programs in continuation-passing style

Plotkin's _v -calculus for call-by-value programs is weaker than the -calculus for the same programs in continuation-passing style (CPS). To identify the call-by-value axioms that correspond to on CPS terms, we define a new CPS transformation and an inverse mapping, both of which are interesting in their own right. Using the new CPS transformation, we determine the precise language of CPS terms closed under -transformations, as well as the call-by-value axioms that correspond to the so-called administrative -reductions on CPS terms. Using the inverse mapping, we map the remaining and equalities on CPS terms to axioms on call-by-value terms. On the pure (constant free) set of -terms, the resulting set of axioms is equivalent to Moggi's computational -calculus. If the call-by-value language includes the control operatorsabort andcall-with-current-continuation, the axioms are equivalent to an extension of Felleisenet al.'s _v -C-calculus and to the equational subtheory of Talcott's logic IOCC.This article is a revised and extended version of the conference paper with the same title [42]. The technical report of the same title contains additional material.The authors were supported in part by NSF grant CCR 89-17022 and by Texas ATP grant 91-003604014.     

Improving Generalization with Active Learning

Active learning differs from learning from examples in that the learning algorithm assumes at least some control over what part of the input domain it receives information about. In some situations, active learning is provably more powerful than learning from examples alone, giving better generalization for a fixed number of training examples.In this article, we consider the problem of learning a binary concept in the absence of noise. We describe a formalism for active concept learning calledselective sampling and show how it may be approximately implemented by a neural network. In selective sampling, a learner receives distribution information from the environment and queries an oracle on parts of the domain it considers useful. We test our implementation, called anSG-network, on three domains and observe significant improvement in generalization.A preliminary version of this article appears as Cohn et al. (1990).     

From “website” to “on-line presence”: From “internet” to “information society”

This paper discusses some of the key drivers that will enable businesses to operate effectively on-line, and looks at how the notion of website will become one of an on-line presence which will support the main activities of an organisation. This is placed in the context of the development of the information society which will allow individuals-as consumers or employees-quick, inexpensive and on-demand access to vast quantities of entertainment, services and information. The paper draws on an example of these developments in Australasia.     

Recognition of handwritten musical notes by a modified Neocognitron

A neural network for recognition of handwritten musical notes, based on the well-known Neocognitron model, is described. The Neocognitron has been used for the what pathway (symbol recognition), while contextual knowledge has been applied for the where (symbol placement). This way, we benefit from dividing the process for dealing with this complicated recognition task. Also, different degrees of intrusiveness in learning have been incorporated in the same network: More intrusive supervised learning has been implemented in the lower neuron layers and less intrusive in the upper one. This way, the network adapts itself to the handwriting of the user. The network consists of a 13×49 input layer and three pairs of simple and complex neuron layers. It has been trained to recognize 20 symbols of unconnected notes on a musical staff and was tested with a set of unlearned input notes. Its recognition rate for the individual unseen notes was up to 93%, averaging 80% for all categories. These preliminary results indicate that a modified Neocognitron could be a good candidate for identification of handwritten musical notes.     

The role of “craft language” in learning “Waza”

The role of craft language in the process of teaching (learning) Waza (skill) will be discussed from the perspective of human intelligence.It may be said that the ultimate goal of learning Waza in any Japanese traditional performance is not the perfect reproduction of the teaching (learning) process of Waza. In fact, a special metaphorical language (craft language) is used, which has the effect of encouraging the learner to activate his creative imagination. It is through this activity that the he learns his own habitus (Kata).It is suggested that, in considering the difference of function between natural human intelligence and artificial intelligence, attention should be paid to the imaginative activity of the learner as being an essential factor for mastering Kata.This article is a modified English version of Chapter 5 of my bookWaza kara shiru (Learning from Skill), Tokyo University Press, 1987, pp. 93–105.     

Choosing reference classes and building provisional models

A well-known problem in default logic is the ability of naive reasoners to explain bothg and ¬g from a set of observations. This problem is treated in at least two different ways within that camp.One approach is examination of the various explanations and choosing among them on the basis of various explanation comparators. A typical comparator is choosing the explanation that depends on the most specific observation, similar to the notion of narrowest reference class.Others examine default extensions of the observations and choose whatever is true in any extension, or what is true in all extensions or what is true in preferred extensions. Default extensions are sometimes thought of as acceptable models of the world that are discarded as more knowledge becomes available.We argue that the notions of specificity and extension lack clear semantics. Furthermore, we show that the problems these ideas were supposed to solve can be handled easily within a probabilistic framework.