Modular design patterns for hybrid learning and reasoning systems

The unification of statistical (data-driven) and symbolic (knowledge-driven) methods is widely recognized as one of the key challenges of modern AI. Recent years have seen a large number of publications on such hybrid neuro-symbolic AI systems. That rapidly growing literature is highly diverse, mostly empirical, and is lacking a unifying view of the large variety of these hybrid systems. In this paper, we analyze a large body of recent literature and we propose a set of modular design patterns for such hybrid, neuro-symbolic systems. We are able to describe the architecture of a very large number of hybrid systems by composing only a small set of elementary patterns as building blocks. The main contributions of this paper are: 1) a taxonomically organised vocabulary to describe both processes and data structures used in hybrid systems; 2) a set of 15+ design patterns for hybrid AI systems organized in a set of elementary patterns and a set of compositional patterns; 3) an application of these design patterns in two realistic use-cases for hybrid AI systems. Our patterns reveal similarities between systems that were not recognized until now. Finally, our design patterns extend and refine Kautz’s earlier attempt at categorizing neuro-symbolic architectures.

     

Fuzzy hypercubes: Linguistic learning/reasoning systems for intelligent control and identification

This paper introduces a new tool for intelligent control and identification. A robust and reliable learning and reasoning mechanism is addressed based upon fuzzy set theory and fuzzy associative memories. The mechanism storesa priori an initial knowledge base via approximate learning and utilizes this information for identification and control via fuzzy inferencing. This architecture parallels a well-known scheme in which memory implicative rules are stored disjunctively. We call this process afuzzy hypercube. Fuzzy hypercubes can be applied to a class of complex and highly nonlinear systems which suffer from vagueness uncertainty and incomplete information such as fuzziness and ignorance. Evidential aspects of a fuzzy hypercube are treated to assess the degree of certainty or reliability. The implementation issue using fuzzy hypercubes is raised, and finally, a fuzzy hypercube is applied to fuzzy linguistic control.     

Constraint reasoning with learning automata

This article presents a decision-maker model, called learning automaton, exhibiting adaptive behavior in highly uncertain stochastic environments. This learning model is used in solving constraint satisfaction problems (CSPs) by a procedure that can be viewed as hill climbing in probability space. the use of a fast learning algorithm that relaxes previous common assumptions is investigated. It is proven that the algorithm converges with probability 1 to a solution of the CSP and a set of test problems show that good performance can be achieved. In particular, it is shown that this method achieves a higher level of performance than that presented in a previous similar approach. Finally, it is estimated the speedup of a parallel implementation and the proposed algorithm is compared with a backtracking algorithm enhanced with standard CSP techniques. © 1994 John Wiley & Sons, Inc.     

Deductive approximate reasoning systems

A formal deductive view for the theory of approximate reasoning, called AR-1, is introduced. A central feature of this framework is the view of propositions as statements involving the assignment of possible values to variables. A unified method for managing joint variables is given. AR-2, which allows for the introduction of probability theory into approximate reasoning, is presented. AR-5, a restrictive version of approximate reasoning, is also introduced     

Formal reasoning in intelligent database systems

Expert database systems were proposed to solve the difficulties encountered in traditional database systems. Prolog provides a fast prototyping tool for building such database systems. However, an intelligent database system implemented in Prolog faces a major restriction that only Horn rules are allowed in the knowledge base. We propose a theorem prover which can make inference for non-Horn intelligent database systems. Conclusions can be deduced from the facts and rules stored in a knowledge base. For a knowledge base with a finite domain, the prover can provide correct answers to queries, derive logical consequences of the database, and provide help in detecting inconsistencies or locating bugs in the database. The theorem prover is efficient in deriving conclusions from large knowledge bases which might swamp most of the other deductive systems. The theorem prover is also useful in solving heuristically the satisfiability problem related to a database with an infinite domain. A truth maintenance mechanism is provided to help eliminate repetitious work for the same goals.Supported by National Science Council under grant NSC 81-0408-E-110-9.     

Internal consistency in plausible reasoning systems

The usual approach to plausible reasoning is to associate a validity measure with each fact or rule, and to compute from these a validity measure for any deduction that is made. This approach is shown to be inappropriate for some classes of problems, particularly those in which the evidence is not internally consistent. Three current plausible reasoning architectures are summarised and each applied to the same small task. An analysis of the performance of these systems reveals deficiencies in each case. The paper then outlines a new approach based on the discovery of consistent subsets of the given evidence. This system can be used either in isolation or in conjunction with a validity-propagating architecture. Comparative results from implementations of all four systems are presented.     

Preference-based reasoning in BDI agent systems

An important feature of BDI agent systems is number of different ways in which an agent can achieve its goals. The choice of means to achieve the goal in made by the system at run time, depending on contextual information that is not available in advance. In this article, we explore ways that the user of an agent system can specify preferences which can be incorporated into the BDI execution process and used to guide the choices made. For example, a user of a travel system can specify a preferred airline, or a particular kind of accommodation, and the system will use this information to satisfy the goal and preferences, if possible. Preferences are specified in terms of properties of goals and resource usage, and are used to make two types of decisions: (a) select a plan when there is a choice and (b) determine the order in which subgoals of a plan should be pursued when their order is not fixed by design. We have implemented our preference framework in Jadex, and provide detailed case studies within the context of a holiday travel agent application.     

From machine learning to machine reasoning

A plausible definition of “reasoning” could be “algebraically manipulating previously acquired knowledge in order to answer a new question”. This definition covers first-order logical inference or probabilistic inference. It also includes much simpler manipulations commonly used to build large learning systems. For instance, we can build an optical character recognition system by first training a character segmenter, an isolated character recognizer, and a language model, using appropriate labelled training sets. Adequately concatenating these modules and fine tuning the resulting system can be viewed as an algebraic operation in a space of models. The resulting model answers a new question, that is, converting the image of a text page into a computer readable text. This observation suggests a conceptual continuity between algebraically rich inference systems, such as logical or probabilistic inference, and simple manipulations, such as the mere concatenation of trainable learning systems. Therefore, instead of trying to bridge the gap between machine learning systems and sophisticated “all-purpose” inference mechanisms, we can instead algebraically enrich the set of manipulations applicable to training systems, and build reasoning capabilities from the ground up.     

CDUL: Class directed unsupervised learning

A novel neural network called Class Directed Unsupervised Learning (CDUL) is introduced. The architecture, based on a Kohonen self-organising network, uses additional input nodes to feed class knowledge to the network during training, in order to optimise the final positioning of Kohonen nodes in feature space. The structure and training of CDUL networks is detailed, showing that (a) networks cannot suffer from the problem of single Kohonen nodes being trained by vectors of more than one class, (b) the number of Kohonen nodes necessary to represent the classes is found during training, and (c) the number of training set passes CDUL requires is low in comparison to similar networks. CDUL is subsequently applied to the classification of chemical excipients from Near Infrared (NIR) reflectance spectra, and its performance compared with three other unsupervised paradigms. The results thereby obtained demonstrate a superior performance which remains relatively constant through a wide range of network parameters.     

Scenario-based knowledge representation in case-based reasoning systems

A scenario-based representation model for cases in the domain of managerial decision-making is proposed. The scenarios in narrative texts are converted to scenario units of knowledge organization. The elements and structure of the scenario unit are defined. The scenario units can be linked together or coupled with others. Compared with traditional case representation methods based on database tables or frames, the proposed model is able to represent knowledge in the domain of managerial decision-making at a much deeper level and provide much more support for case-based systems employed in business decision-making.     

Cooperative learning neural network output feedback control of uncertain nonlinear multi-agent systems under directed topologies

Without assuming that the communication topologies among the neural network (NN) weights are to be undirected and the states of each agent are measurable, the cooperative learning NN output feedback control is addressed for uncertain nonlinear multi-agent systems with identical structures in strict-feedback form. By establishing directed communication topologies among NN weights to share their learned knowledge, NNs with cooperative learning laws are employed to identify the uncertainties. By designing NN-based κ-filter observers to estimate the unmeasurable states, a new cooperative learning output feedback control scheme is proposed to guarantee that the system outputs can track nonidentical reference signals with bounded tracking errors. A simulation example is given to demonstrate the effectiveness of the theoretical results.     

A simple learning environment improves mathematical reasoning

Abstract

ANIMATE, an interactive computer animation-based tutor, has been developed as part of an on-going test of a theory of word problem comprehension. Tutor feedback is unobtrusive and interpretive: unexpected behavior in the equation-driven animation of a situation highlights equation errors which the student resolves through iterative debugging. The student has responsibility for learning, goal-setting and diagnosis. Experimental controls (n = 96) with Motion problems show that improvement cannot be solely attributed to practice, computer use, or use of the situation-based method. Concurrent think-aloud protocols of students (n = 7) solving Motion, Work and Investment problems over two days (in a pretest-posttest design) uncover specific changes which underlie these improvements. ANIMATE is an effective problem-solving aid, and there is transfer of learning. Problems with impossible situations were acknowledged by median-level subjects (posttest scores between 77% and 85%), but solved blindly by high-level subjects (post test scores > = 95%), suggesting an automatically controlled processing dichotomy. On Day 2, subjects spent more time reviewing problem texts and correcting flawed expressions. They developed self-directed debugging skills, reminiscent of expert problem-solving in many domains, without relying on tutor feedback behaviors. The system is unintelligent by ITS standards but communicates knowledge to the students, helping them teach themselves approaches for mathematical problem-solving.     

Heterogeneous relational reasoning in knowledge graphs with reinforcement learning

Path-based relational reasoning over knowledge graphs has become increasingly popular due to a variety of downstream applications such as question answering in dialogue systems, fact prediction, and recommendation systems. In recent years, reinforcement learning (RL) based solutions for knowledge graphs have been demonstrated to be more interpretable and explainable than other deep learning models. However, the current solutions still struggle with performance issues due to incomplete state representations and large action spaces for the RL agent. We address these problems by developing HRRL (Heterogeneous Relational reasoning with Reinforcement Learning), a type-enhanced RL agent that utilizes the local heterogeneous neighborhood information for efficient path-based reasoning over knowledge graphs. HRRL improves the state representation using a graph neural network (GNN) for encoding the neighborhood information and utilizes entity type information for pruning the action space. Extensive experiments on real-world datasets show that HRRL outperforms state-of-the-art RL methods and discovers more novel paths during the training procedure, demonstrating the explorative power of our method.     

Student-teacher communication directed to computer-based learning environments

The main objective of this paper is to provide an optimal way for students to control a computer-based teacher and interact with it. From previous experience, we took a special interest in improving the navigation through tutorial systems, and also in the evaluation techniques, designing and developing a new tutorial based on an open architecture. After implementing the tutorial prototype, it was evaluated by a selected group of users in a controlled laboratory situation in order to gather data about the characteristics and usability of this prototype. In general, users had a good overall opinion of the evaluated tutorial. The key idea behind our experience is the introduction of tutorials in all practical classes as a complement to the instructor in the near future.     

Object-oriented representations, causal reasoning and expert systems

Abstract: In this article we describe how two popular AI representation techniques—frames and production systems— can be usefully combined under a general AI object-oriented approach to problems which arise in the domain of hardware fault diagnosis. One of the main advantages of such a combination is that causal reasoning— crucial to the domain under consideration — can also be naturally and effectively represented. In order to put the combination of frames and production systems on a sound methodological footing, we first provide a knowledge engineering methodology which is an object-oriented re-interpretation of ontological analysis.