Cognitive and Meta-Cognitive Agents

Cognition refers to the process of human thought. It involves understanding how the mind works. In research, there were significant efforts in cognition research shaping different education theories. On the aspect of technology, there were several efforts to build cognitive models with agents. Scholars such as Allan Newell laid the foundation to unified cognition theories that have been used later with agents creating SOAR architecture. Steve, the pedagogical agent (see Section 4.1) was created based on SOAR. Another popular cognitive system is the ACT-R architecture developed at Carnegie Mellon University. Dubois, Nkambou, Quintal, and Savard (2010) provide an introduction, history, and review of cognitive architectures with inspiration of their biological roots. Dubois et al.

(2010) argue on the rationale and the necessity of understanding human cognition to provide effective tutoring systems. Not only understanding the human cognition can lead to better learning, but also to building artificially intelligent entities that simulate the human brain as possible, such as the case of Steve with SOAR stated above. In this case, it becomes possible to teach an artificial agent harvesting benefits of learning by teaching². However, fully understanding human cognition and building a complete model is far from reaching. Yet, researchers are working to find better representing models day after day.

Cognitive models can further help in achieving co-cognition between the pedagogical agent and the learner. Dubois, Nkambou, Quintal, and Savard (2010) point the challenges researchers face with this understanding. Researchers further attempted employing cognitive architectures to also consider emotions, as discussed later in this Chapter. Cognitive tutors are agents with cognitive architectures mainly adopted in Intelligent Tutoring Systems. ACT-R is an example cognitive tutor agent that is based on the ACT-R cognitive theory.

Furthermore, researchers have pointed the importance of building architecture of cognitions rather than a single entity. Franklin and Graesser (2001) indicated that the agent model is a suitable model for building a cognitive architecture, and hence provided the Intelligent Distribution Agent (IDA) based

1 Several paragraphs of this section are adopted from Soliman and Guetl (2011c).

2 It is found in research that it is possible to teach an agent by the exchange of a concept map. Teachable agents are discussed later in this Chapter.

on the so called the Global Workspace cognitive architecture (GW)¹. Thus the cognition based approach gives also a rise to intelligent agents. Dubois, Nkambou, Quintal, and Savard (2010) propose the following definition of an ideal artificial cognitive tutoring agent: “an agent built on an architecture that offers structures, features and functioning comparable to the human model so that it is similarly capable of adaptation, learning, generalization within and across domains, and action in complex situations encountered in tutoring learners.”

Understanding how knowledge is constructed will lead to designing effective teaching methods. Relating to prior knowledge by a schema of prior concepts has an influence on adaptive systems that present tailored instruction. The critical thinking methods compliant with cognitive theories can lead to agent design methods with an agent providing contradictions stimulating thinking and leading to better established knowledge. Thus it is also important to consider cognitive disequilibrium as well.

Conceptual Change Agents

Conceptual change is an important and evolving learning approach especially to science education. Conceptual change involves deep changes of one’s knowledge. And therefore concepts and their relationships change over time. It can involve being exposed to contradictions and critical thinking to reach cognitive equilibrium according to Jean Piaget constructivism theories. This equilibrium can be exposed to new set of conflicts through conceptual change leading to a new cognitive equilibrium that is a better established knowledge.

Ting and Chong (2003) propose the integration of conceptual change in animated pedagogical agents. Learners perform experiments and are given an opportunity to generalize or model the experiment by linking graph nodes of concepts thus creating hypotheses. The pedagogical agent will foster conflicts and thus stimulate the conceptual change for the learners. The intelligent agent also has the function of changing its conceptualization by the so called belief revision functions in the Belief-Desire-Intention (BDI) model that will be discussed later.

While learning is a social activity, social interactions have a learning component. I.e., one may learn by social interactions with others by influencing each other concepts, possibly discussions and arguments are triggers for concept change leading to a new equilibrium. Distributed cognition vs. collective cognition looks at surrounding objects and group of individuals rather than individualized learning ones. Study of this model is relevant to multi-agents and the social virtual environment.

1The GW theory focuses on consciousness and unconsciousness in learning.

5.3.1. Bayesian Networks for Agent Learning

Bayesian Networks (BN) is a common probabilistic AI method. It is common to be referred also as Belief Network (Russel & Norvig, 1995). A BN is represented as a directed acyclic graph with each node represents a belief or hypothesis as a random variable while edges represent conditional dependences between them. The BN learning problem is inferring the structure of the graph from data, which is not a trivial problem. Consequently, BNs are used for inferences as a method for inferring about learner abilities to provide tailored adaptive instruction consequently. The main interest is the machine learning application and issues relevant to understanding the learners and their cognition and dynamically dealing with it through time. A Decision Network (DN) is an extension to the Bayesian Network by adding utility and decision nodes to enable solving decision problems under uncertainty (Russel & Norvig, 1995, pages 471, 484). To handle evolution of the state of the environment over time, Dynamic Belief Networks (DBN) and Dynamic Decision Networks (DDNs) are devised as variations to BN and DN accordingly to enable further representation changes and solve complex decisions problems (Russel & Norvig, 1995, pages 514, 516).

The Dynamic Bayesian Network (DBN) model has been used by Conati and Zhao (2004) to enable a learner model through a pedagogical agent. The choice of the pedagogical agent was due to an evidence of increased learner engagement and improved learning as discussed in Chapter 4. The learner model by Conati and Zhao (2004) considers cognitive, meta-cognitive abilities, and emotions.

Upon constructing the DBN, reasoning about learner knowledge levels is done and used by the pedagogical agent to provide guidance to the learner. By having more information through this model about the affective state of the learner, the purpose of that study is to have better informed pedagogical agent interventions.

The work by Conati and Zhano (2004) provides relevant details such as: what DBN to keep, short term, long term student models, and an evaluation of the results. Ting and Chong (2006) equip pedagogical agents within the scientific inquiry learning environments (INQPRO) with DBNs. In this model, a probabilistic BN model is used to model learner properties as it changes through time due to learning or conceptual change mentioned above. Learners interact with the learning lesson as a computer simulation by the aid of a pedagogical agent. At this stage, a mental model is constructed. Then a discrepancy is presented through the simulation. The pedagogical agent monitors the subsequent interactions through this process of conceptual change leading to conflict resolution stage. In this work, the probabilistic nature of capturing mental states is tackled by the use of Bayesian Networks and the conceptual change mandated it to be of a dynamic nature. Ting and Chong (2006) did not consider emotions nor checked for the accuracy of what is captured in the above mentioned work.

The DDN concept has been used in several tutoring systems for predicting learning abilities such as knowledge, focus of attention, affective states, and actions taken. The purpose of the DDN according to the research by Ting and Phon-Amnuaisuk (2010) is to support scientific inquiry by reasoning about learners abilities and provide learning support. A special feature of the DDN is its ability to deal with the temporal aspects captured during interaction with the system (Ting & Phon-Amnuaisuk, 2010). Peedy, the pedagogical agent (see Table 6), uses DDN in the INQPRO learning environment.

Common to the above mentioned research efforts with BN, DDN, and conceptual change models are trying to capture concepts or cognitive states of the learner by observing the behavior of interaction with the system.

5.3.2. BDI Agents for Human Learning

The Belief-Desire-Intention model (BDI) has been widely adopted in agent-based systems. It originates from AI Agent research (Jennings & Wooldridge, 2000) as it encodes agent behavior though setting goals that determine desires and the intentions. Generally, the agent behavior can result from the BDI model (Weiss, 1999, p. 54). The BDI model is composed of a set of logic rules to form the cognitive encoding of the agent as a core component. The components of the BDI model are explained in Table 8 below.

Table 8: Belief-Desire-Intention (BDI) model (Jaques & Viccari, 2004).

Beliefs “represent the information about the state of the environment that is updated appropriately after each sensing action.”

Desires “are the motivational state of the system. They have information about the objectives to be accomplished.”

Intention “is a desire that was chosen to be executed by a plan, because it can be carried out according to the agent’s beliefs (because it is not rational that the agent carries out something that it does not believe). Plans are pre-compiled procedures that depend on a set of conditions for being applicable.”

Hence, the BDI model gives autonomy and goal-directed behavior to the intelligent agent as a main distinction between it and a traditional object. The BDI model is used by agents to create behavior by creating to achieving goals.

Those agent types monitor the environment to capture changes and represent them in their beliefs. BDI systems have two main reasoning processes: reasoning about the intentions to undertake and reasoning about the desires and whether they have been reached.

With the reasoning of the BDI model, it can be further useful in the important learning activity of explanation by track changes in the beliefs, desires, and the intentions and showing how reasoning happened. Broekens et al. (2010)

utilize BDI agents for algorithms of explaining the agent behavior to the user. It uses an agent platform, named GOAL¹ that supports the BDI concept for that purpose. This work relates the reasoning chain to a result to show how the agent has reached that state thus achieving an explainable agent for learning or training purpose. This concept is referred as explainable agents who can explain and convey their behavior to the learner or to an educator entity.

Bercht and Vicari (2000) extend cognitive encoding in agents to add the emotional dimension of the learner to the learner profile. The affective domain in the learner is explored through mental states of the learner exploited by the use of BDI-based agents. This allowed researchers to model mental states and their complex interactions allowing describing agent complex activities. Reasoning about student current emotional state is very important in the classroom and in electronic environments as well (see Sections 2.1.4, 4.2). The tutor can utilize those states to improve motivation, restrain about stressors and it may provide stressors in specific situations that can improve learning. Jaques and Viccari (2004) also utilize the inference power of BDI to reason about the student emotions given its dynamic nature. The system needs to know events in the environment, the student’s goals, and the desirability of the events according to student’s goals. Jiang, Vidal, and Huhns (2007) extend the BDI model integrating the emotional dimension to the Emotional BDI (EBDI) architecture. The model describes interactions between beliefs, desires, and intentions with emotions during the reasoning process.

5.3.3. Meta-Cognitive and Self Regulated Learning Agents

Meta-cognition is an important topic to learning. Meta-cognition is relevant to thinking about thinking and is related to learning improvement as it calls for learners’ awareness of their cognitive abilities and learning gains. Critical thinking for example is related to meta-cognitive ability. Self Regulated Learning (SRL) is also a cognitive activity. SRL means learning that is guided by meta-cognition with strategic actions, evaluating, and motivation to learn. Zimmerman (1990) provides detailed overview and definitions of self regulated learning, and shows its impact on student achievement. The importance of this subject is for learners to be self-guided and self-directed having learning goals directing them by having self-control of cognitive processes.

Kinnebrew, Biswas, Sulcer, and Taylor (2011) employ self regulated learning aided by Betty’s pedagogical agent. Betty’s system of SRL is depicted in Figure 14. As shown in the Figure, Betty provides either knowledge construction or

1 GOAL as a multi-agent realization framework will be discussed in Chapter 8.

monitoring strategies. Information structuring meta-cognitive student activity is communicated to Betty’s system through concept maps. Students conduct monitoring meta-cognitive activities with Betty such as asking questions to the pedagogical agent (checking) and probing parts of the concept map with the pedagogical agent (probing). In this research, the pedagogical agent provided meta-cognitive feedback to the learners. While this study provided results showing that students perform better in knowledge construction strategy, it sheds light into the importance of developing monitoring meta-cognitive strategies to improve self-learning. The study reports positive results on learning by teaching a pedagogical agent as well.

Figure 14: Model of self regulated learning strategies and activities in Betty’s Brain agent (Kinnebrew, Biswas, Sulcer, & Taylor, 2011).

Hence, self-guided or self-regulated learning is fundamental to active learning realizations in virtual learning environments that lack of human teacher but being subsidized with artificially intelligent agents. Therefore, one can seek to utilize intelligent agents with SRL and meta-cognition abilities in the realization of pedagogical agents as well to promote self-guided learning in VLE.