Elsevier

Knowledge-Based Systems

Volume 30, June 2012, Pages 48-56
Knowledge-Based Systems

A computational model for causal learning in cognitive agents

https://doi.org/10.1016/j.knosys.2011.09.005Get rights and content

Abstract

To mimic human tutors and provide optimal training, a cognitive tutoring agent should be able to continuously learn from its interactions with learners. An important element that helps a tutor better understand learner’s mistake is finding the causes of the learners’ mistakes. In this paper, we explain how we have designed and integrated a causal learning mechanism in a cognitive agent named CELTS (Conscious Emotional Learning Tutoring System) that assists learners during learning activities. Unlike other works in cognitive agents that used Bayesian Networks to deal with causality, CELTS’s causal learning mechanism is implemented using data mining algorithms that can be used with large amount of data. The integration of a causal learning mechanism within CELTS allows it to predict learners’ mistakes. Experiments showed that the causal learning mechanism help CELTS improve learners’ performance.

Introduction

Causal learning is the process through which we come to infer and memorize an event’s reasons or causes based on previous beliefs and current experience that either confirm or invalidate previous beliefs [1], [2]. Finding the causes of the problems helps human beings to better deal with everyday problems. For instance what are the causes of the global warming? Different methods are proposed for finding causal relations between events such as scientific experiments, statistical relations, temporal order, prior knowledge, and so forth [2], [3]. Researchers in cognitive computer sciences have been interested in causality and how causal learning can be implemented in cognitive agents [4]. For example, researchers simulated causal mechanisms in cognitive architectures such as ACT-R [5] and CLARION [6]. However, these works have one or more of the following limitations:

  • Assumptions about the causal model or values to variables have to be specified by a programmer or domain expert.

  • Causal mechanisms are implemented by using techniques that are not scalable to handle a large amount of data.

  • Mechanisms do not include a role for emotions.

In this paper, we will discuss our attempt for the implementation of causal learning in the Conscious Emotional Learning Tutoring System (CELTS) [7], which addresses these limitations. CELTS is a general cognitive architecture designed to be put to work as a Tutor for astronauts learning to manipulate the International Space Station’s (ISS) robotic telemanipulator, Canadarm2. CELTS architecture relies on the functional “consciousness” [8] mechanism for much of its operations. It also bears some functional similarities with the physiology of the nervous system. Its modules communicate with one another by contributing information to its Working Memory through information codelets1 [9]. In this study, CELTS is integrated in CanadarmTutor [10], [11] (Fig. 2), a simulation-based intelligent tutoring system for learning how to operate the Canadarm2 robotic arm (Fig. 1) installed on the International Space Station (ISS). CanadarmTutor learning environment is a 3D reproduction of Canadarm2 on the space station and its control panel (Fig. 2). Learning activities in CanadarmTutor mainly consists of operating Canadarm2 for performing various real-life tasks with the simulator such as carrying loads with the robotic arm or inspecting the ISS. Operating Canadarm2 is a complex task because astronauts have to follow a strict security protocol. Furthermore, Canadarm2 has seven-degrees of freedom (seven joints that can be rotated) and users only have a partial view of the environment through the cameras that they choose and adjust. CanadarmTutor was the subject of several research projects [10]. CELTS is the component of CanadarmTutor that acts as the core tutor which takes all the pedagogical decisions, generates dialogue and performs the high-level assessment of the learner performance.

The learners’ manipulations of the virtual world simulator (Fig. 2), constitute the interactions between CELTS and its users. In particular, the virtual world simulator sends all manipulation data to CELTS, which in turn sends learners various types of advice to improve their performance (Fig. 3).

We have implemented a general emotional mechanism (emotions and emotional learning), episodic learning2 (EPL) [7] and causal learning [12] in CELTS. In the context of CELTS, we refer to causal learning as the use of inductive reasoning to generalize causal rules from sets of experiences. CELTS observes astronauts’ behavior without complete information regarding the reasons for their behavior.

CELTS’s causal learning in its current implementation is capable of finding the causes of user mistakes. For instance, as an example of the real world, suppose one observes that each time one forgets to adjust his car’s side and front mirrors (M), he tends to have poor control over the wheel (W) and cause collision risk (C) with other cars. We can link these variables in the following way:MWC;WMC.

The first graph (1) shows that the probability of forgetting mirror adjustment is independent of the probability of a collision risk with other cars, but it is in turn conditional on the occurrence of poor wheel control. The second graph (2) demonstrates that the probability of poor wheel control is independent of the probability of making a collision with other cars and is conditional on forgetting mirror adjustment. To give to CELTS the capability of causal learning to learn complex causal relationships like the above example, such that events maybe conditionally dependent or independent, we designed a causal learning mechanism [7]. However, it was found in our experiment that the number of rules learned can be very large for CELTS to handle. Because CELTS is a tutor that interacts with many learners, so, it is faced with a huge amount of data. We observed that at any given time, only a small subset of the rules is relevant for the current situation.

To determine which rule best matches with the current problem, we present in this paper a modified version of the causal learning mechanism described previously [12]. Our modified algorithm relies on a custom data mining algorithm that is designed to efficiently mine only the rules that are relevant to the current situation. To do so, we have changed our previous algorithm to accept a time constraint and other constraints on rules to be mined. As it will be shown in the experimental evaluation, using the constraints (1) can reduce the number of rules found and the execution time of the causal learning algorithm by several orders of magnitude, and (2) the new causal learning algorithm is therefore scalable to handle large amount of data (more than 20,000 sequences) very efficiently. This is an important contribution because in current cognitive agents, causal learning has up to now been implemented with techniques such as Bayesian networks that are not scalable for handling a large amount of data.

The rest of this paper is organized as follows. In Section 2, we will give a brief review of the related works about causal learning in cognitive science. In Section 3, we will briefly explain CELTS’s emotional and episodic learning mechanisms. In Section 4, we propose an improved causal learning mechanism for CELTS. In Section 5, we present results from our experiments with CELTS. Finally, in Section 6, we draw a conclusion.

Section snippets

Related works

Up to now, scientists mainly proposed to use causal Bayes nets (acyclic graphs to establish causal relation between events. The key issue for the construction of a causal Bayes net is finding conditional probabilities between events. Mathematics is used to describe conditional and unconditional probabilities between a graph’s variables. The structure of a causal graph restricts the conditional and unconditional probabilities between the graph’s variables. We can find the restrictions between

CELTS’s architecture

CELTS (Fig. 4) is a cognitive agent architecture based on Baars’ theory [14] of consciousness. It is constructed with simple agents called “codelets” (which reproduce Baars’ “simple processors”). The central point of the system is the “access consciousness”, which allows all resources to access centrally selected information that is “broadcast” to unconscious processes (which guides the agent to be stimulated only with the most relevant information). CELTS performs through cognitive cycles. A

Integrating causal learning in CELTS

We now describe how we have designed and integrated a causal learning mechanism into CELTS architecture.

Testing causal learning in the New CELTS

To determine the extent to which the improved causal learning mechanism improved CELTS’ performance, we asked 8 users to test the new version of CELTS with improved causal learning mechanisms (version A) and the version of CELTS with the previous causal learning mechanism (version B). Learners were invited to manipulate Canadarm2 for approximately 2 h, using both versions A and B of the system. The first four students (group A) used version A, and then version B. The second four learners (group

Conclusion

In this paper, we described how our improved data mining algorithm can improve a cognitive agent to find the cause of the learners’ mistakes. The new and the previous versions of CELTS’ causal learning algorithm were compared in four ways. First, CELTS’ performance was evaluated based on the number of correct interventions given to the learners during training sessions. Second, we measured the impact on learners’ performance. Third, we evaluated the satisfaction of users. Fourth, a domain

References (31)

  • A. Maldonado, A. Catena, J. C. Perales, A. Cándido, Cognitive biases in human causal learning,...
  • B. Daniel et al.

    Psychology of Learning and Motivation Moral Judgment and Decision Making

    (2009)
  • D.A. Lagnado et al.

    Beyond Covariation: Cues to Causal Structure

  • A. Gopnik et al.

    Causal learning: Psychology, Philosophy, and Computation

    (2007)
  • W. Schoppek, Stochastic independence between recognition and completion of spatial patterns as a function of causal...
  • S. Hélie, Modélisation de l’apprentissage ascendant des connaissances explicites dans une architecture cognitive...
  • U. Faghihi, The use of emotions in the implementation of various types of learning in a cognitive agent, Ph.D, Computer...
  • S. Franklin et al.

    The LIDA Architecture:Adding New Modes of Learning to an Intelligent, Autonomous, Software Agent

    (2006)
  • D.R. Hofstadter, M. Mitchell, The Copycat Project: A model of mental fluidity and analogy-making, in: K.J. Holyoak,...
  • R. Nkambou, K. Belghith, F. Kabanza, An approach to intelligent training on a robotic simulator using an innovative...
  • R. Nkambou et al.

    Learning task models in ill-defined domain using an hybrid knowledge discovery framework

    (2011)
  • U. Faghihi, P. Fournier-Viger, R. Nkambou, P. Poirier, A Generic causal learning Model for Cognitive Agent, in: The...
  • M. Braun, W. Rosenstiel, K.-D. Schubert, Comparison of Bayesian networks and data mining for coverage directed...
  • B.J. Baars

    In the Theater of Consciousness:The Workspace of the Mind

    (1997)
  • P. Maes

    How to do the right thing

    Connection Science

    (1989)
  • Cited by (0)

    View full text