introduction to the carina metacognitive architecture
TRANSCRIPT
See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/328151295
Introduction to the CARINA Metacognitive Architecture
Conference Paper · July 2018
DOI: 10.1109/ICCI-CC.2018.8482051
CITATIONS
0READS
168
4 authors, including:
Some of the authors of this publication are also working on these related projects:
Analysis of verbal protocols in the resolution of programming logic problems in sixth grade students of the Mercedes Abrego educational institution View project
CARINA: Cognitive architecture for reasoning failure detection in Intelligent Agents (Focused in Intelligent Tutoring Systems) View project
Manuel Fernando Caro Piñeres
Universidad de Córdoba (Colombia)
40 PUBLICATIONS 58 CITATIONS
SEE PROFILE
Adan Gomez
Universidad de Córdoba (Colombia)
16 PUBLICATIONS 3 CITATIONS
SEE PROFILE
Catriona M Kennedy
University of Birmingham
44 PUBLICATIONS 352 CITATIONS
SEE PROFILE
All content following this page was uploaded by Adan Gomez on 12 October 2018.
The user has requested enhancement of the downloaded file.
Introduction to the CARINA metacognitive architecture
Manuel F. Caro Dept. Informática Educativa
Universidad de CórdobaMontería, Colombia
Darsana P. JosyulaDepartment of Computer
Science, Bowie State University, Bowie, MD – USA
Adán A. Gómez Dept. Informática Educativa
Úniversidad de Córdoba Montería,Colombia
Catriona M. Kennedy School of Computer Science,
University of Manchester, Manchester, UK
Abstract— Metacognition has been used in artificial intelligence to increase the level of autonomy of intelligent systems. However the design of systems with metacognitive capabilities is a difficult task due to the number and complexity of processes involved. This paper presents the CARINA architecture, which is based on precise definitions of structural and functional elements of metacognition as defined in the MISM metamodel. CARINA can be used to implement real-world cognitive agents with the capability for introspective monitoring and meta-level control. Introspective monitoring detects reasoning failure (for example, when expectation are violated). Metacognitive control selects strategies to recover from failures. The paper demonstrates a CARINA implementation of reasoning failure detection and recovery in an intelligent tutoring system called FUNPRO. The tutoring system also searches for possible explanations of a failure by searching for known explanations and by analyzing its reasoning trace.
Keywords—metacognition, cognitive architecture, cognitive computing, CARINA
I. INTRODUCTION
Computational metacognition is an area of Artificial Intelligence (AI) that focuses on the ability of intelligent systems to monitor and to control their own learning and reasoning processes [1]–[3]. Computational metacognition distinguishes reasoning about reasoning from reasoning about the world [4].
A metacognitive architecture provides a concrete framework for detailed modeling of mechanisms for an AI agent's high-level reasoning about itself, through specifying essential structures, divisions of modules, relations among modules, and a variety of other essential aspects [5], [6].
Metacognitive architectures differ from cognitive architectures in that the agent itself is the referent of the cognitive processing but share a commitment to formalisms for representing knowledge, memories for storing this domain content, and processes that utilize and acquire the knowledge [6], [7].
Nowadays increasingly complex AI agents that make decisions based on multiple variables are developed. The complexity increases the probability that reasoning failures occur in such AI agent [8], [9]. A reasoning failure is defined as an outcome other than what is expected or a lack of some outcome or appropriate expectation [10]. Reasoning failures are generated mainly by unfinished tasks or unexpected results in the performance of a task. Metacognition provides introspective monitoring and meta-level control mechanisms for an AI agent to detect and correct its own reasoning failures.
Some implementations exist and have made progress in the integration of metacognitive functions in cognitive architectures. The ACT-R [11]–[13] is a cognitive architecture based on the theory of rational analysis can provide a description of the processes from perception through to action for a wide range of cognitive tasks [13]. Although ACT-R theory was not developed with specific metacognitive mechanisms, several metacognition research has been developed based on this architecture. For example, the work of [14] demonstrate through the use of ACT-R that subjects can dynamically control their problem solving process based on the knowledge of the relative demands of tasks and the mental resources needed to complete it. Reitter [15] designed several metacognitive layers in ACT-R that implement several strategies to address the basic control task, as well as the means to classify and select those strategies according to their suitability in a given situation. However, the modeling process has concentrated on multitasking behavior and has not clarified the specific relationship between multitasking and metacognition .
Research on metacognitive architectures includes the CLARION architecture [5] which allows designers to construct models of specific metacognitive processes, which are then used to capture experimental data related to meta-cognition using a a meta-cognitive subsystem. The meta-cognitive subsystem (the MCS) monitors and controls/regulates cognitive processes for the sake of improving cognitive performance in various circumstances [5], [16]. The role of the meta-cognitive subsystem is to monitor, direct, and modify the operations of all the other subsystems [16]. Actions in the
Proc. 2018 IEEE 17th Int’l Conf. on Cognitive Informatics & Cognitive Computing (ICCI*CC’18) Y. Wang, S. Kwong, J. Feldman, N. Howard, P. Sheu & B. Widrow (Eds.)978-1-5386-3360-1/18/$31.00 ©2018 IEEE
meta-cognitive subsystem can be in form of setting goals, setting parameters and changing an on-going process in other subsystems [5], [16]. However, Clarion does not present processes for the explanation of the cause of a possible poor cognitive performance.
MIDCA is other metacognitive architecture, which is based on earlier work on computational metacognition including the MetaCognitive Loop [17], [18] and Introspective Multistrategy Learning theory [19]. The MIDCA architecture consists of “action-perception” cycles at both the cognitive (i.e. object) level and the metacognitive (i.e., meta-) level (cox). MIDCA is focused on providing agent with greater autonomy in solving problems. Goal insertion is a fundamental process in MIDCA and occurs at both the object level and meta-level. At the meta-level, the cycle achieves goals that change the object level and metacognitive “perception” components introspectively monitor the processes and mental state changes at the cognitive level. MIDCA and CLARION do not present mechanisms of software engineering to simplify the process of developing a intelligent software agent with integration of metacognitive capabilities.
In the context described, the main objective of this paper is to introduce a novel metacognitive architecture for monitoring and control of reasoning failures in artificial intelligent agents. The metacognitive architecture is named CARINA and has support for introspective monitoring and meta-level control. CARINA attempts a thorough integration of action and perception with both cognition and metacognition. These kind of reasoning systems are called “perpetual self-aware cognitive
agents” [20], [21].
CARINA allows cognitive systems to detect reasoning failure, to explain the failure and to select stratetegies for recovery. The implementation of CARINA aims to simplify the process of developing a intelligent software agent and to permit the user to focus on the specifics of the application, hiding the complexities of the generic parts of the model.
The implementation of CARINA aims to simplify the process of developing a intelligent software agent and to permit the user to focus on the specifics of the application, hiding the complexities of the generic parts of the model
In this paper, first the CARINA architecture is sketched. Next, a domain-specific visual language specifically developed for modeling metacognition in intelligent systems called M++ is described. Finally, a metacognitive model and a computational implementation based on CARINA is presented.
II. THE CARINA ARCHITECTURE
CARINA is a metacognitive architecture for artificial intelligent agents. CARINA is derived from the MISM Metacognitive Metamodel [22]. CARINA integrates self-regulation and metamemory with support for the metacognitive mechanisms of introspective monitoring and meta-level control; in this sense CARINA assumes a functional approach to philosophy of mind, according to [23]–[25]. According to [26], the term ‘‘mechanism” includes entities and activities involving entities (i.e., including both static and dynamic
aspects). The entitities in CARINA are called “cognitive elements”.
In accordance with MISM, CARINA is composed of three types of cognitive elements: structural elements, functional
elements and basic elements. Structural elements are containers into which the functional and basic elements are embedded; the main structural element is the “Cognitive
level”. The functional elements are tasks that enable reasoning and decision-making. Basic elements consist of the set of elements that participate and interact in reasoning and meta-reasoning processes. The main functional elements in CARINA are: reasoning task and meta-reasoning task.Reasoning tasks (RT) are actions that enable the processing (transformation, reduction, elaboration, storage and retrieval) of information by applying knowledge and decision making in order to meet the objectives of the system. A metareasoning task (MT) may be used to explain errors in some reasoning task or it may be used to select between “cognitive algorithms” to perform the reasoning [2].
A. Cognitive Loop in CARINA
CARINA is composed of two cognitive levels named object-level and meta-level. The object-level contains the model that an artificial intelligent agent has for reasoning about the world (i.e. agent’s environment) to solve problems [22]. The meta-level is a level of representation of the reasoning of an artificial intelligent agent. The meta-level includes the components, knowledge and mechanisms necessary for a system to monitor and control its own learning and reasoning processes.
1) Cognitive Loop at the Objetct-Level in CARINA
A cognitive loop (action-perception cycle) in CARINA starts when the agent perceives changes to the environment resulting from actions. The situation assessment stage takes as input the perception and processes it. The processing of the information perceived includes the recognition of situations or events as instances of known or familiar patterns and the categorization of objects, situations, and events into known concepts or categories. The output of the situation assessment stage is the combination of perceptual information about many objects and events to compose a comprehensive model of the current environment (i.e. model of the world).
Fig. 1. General view of Object-level in CARINA
The reasoning process in CARINA takes as input beliefs, assumptions and expectations. It then reaches some mental conclusion (e.g. categorization). This is a central cognitive activity that enables an agent to augment its knowledge state. The problem solving process is closely related to reasoning in the sense that it is given a goal (in the world or within the agent) and the agent then chooses and orders its actions to achieve this goal. Monitoring a plan’s execution can also lead to revised estimates about the plan’s effectiveness, and, ultimately, to a decision to pursue some other course of action. In CARINA, decisions are often associated with the recognition of a situation or pattern, and combines the two mechanisms in a recognize-act cycle that underlies all cognitive behavior. Finally, acting involves the represention and storage of motor skills that enable the activities stored in Procedural Memory.
2) Metacognitive Loop at the Meta-Level in CARINA
The meta-level keeps an updated model of the object-level called the “self-model”. This self-model is based on an internal representation of the reasoning processes that occur at the object-level. The self-model stores the data generated by the reasoning processes, along with the rules that were used to address the reasoning. The self-model allows the interchange of information between the meta-level and the object-level. The self-model enables an intelligent system to reason about its own reasoning, since it describes the internal state of the system. The self-model allows the meta-level to have awareness about reasoning processes that are conducted at the object-level.Introspective monitoring and meta-level control are two meta-reasoning mechanisms implemented at the
meta-level in CARINA. Using the common terms of cognitive science, the notion of ‘‘mechanism”involves both representations as well as cognitive mechanisms and processes operating on them [27]. Introspective monitoring includes mechanisms for detecting reasoning failures at the object-level. The main purpose of monitoring is to provide enough information to make effective decisions in the meta-level control. The monitoring process is done through information feedback that is gathered at the meta-level from the object level, see Fig. 2. Thus each cognitive task executed in the object level has a performance profile that is continuously updated in the meta-level. The performance profile is used to evaluate the results of each reasoning task. For example, a robot explores a building and uses its current map of each floor by predicting what kind of room it will find at each location. The performance profile summarizes how well the reasoning is performing and may include its prediction accuracy of what it finds when it opens a door.
Fig. 2. Specification of meta-reasoning mechanisms in
CARINA
ProfileGeneration, FailureDetection, FailureExplanation and GoalGeneration are the main monitoring tasks in CARINA: The ProfileGeneration task performs the interface between the meta-level and the object-level.
When a cognitive task is running, it generates computational data. ProfileGeneration reads the computational data and generates a Profile of the CognitiveTask. Each CognitiveTask in the object-level has a performance Profile in the meta-level; thus the meta-level is always informed of the status of the reasoning in the object-level.
The Sensor has the function of monitoring the profiles of cognitive tasks in order to detect disturbances or anomalies that may represent reasoning failures produced by the cognitive task. FailureDetection reads the properties of a Sensor. If the Sensor finds a discrepancy between observations and expectations regarding the performance of the CognitiveTask, then FailureDetection detects a ReasoningFailure in the CognitiveTask being monitored. FailureExplanation generates an
Explanation of the cause of the ReasoningFailurehaving as inputs the assessment of the failure and reading the ReasoningTrace. GoalGeneration produces new goals based on the Explanation for solving the failure detected. A plan to solve the ReasoningFailure is built the basis of the new Goal. The plan is called FailureSolutionPlan. In the case of the robot, an explanation might be that its map is incorrect and it needs to add more details to it (new goal). For example, this building might be an exception to the norm on which its map is based. Fig. 2 shows the specification of the introspective monitoring process in CARINA.
On the other hand, the meta-level control function decides whether to invoke a scheduler, which scheduler to invoke, and the quantity of resources to invest in the scheduling process [28]. The meta-level control in CARINA includes rules for the recommendation of the best strategy available to correct some reasoning failure at the object-level. The main objective of the meta-level control is to improve the quality of decisions made by the system. Metacognitive tasks associated with meta-level control may explain errors in the reasoning task or it may select between cognitive “algorithms” to perform the reasoning [4].
Therefore, the main control tasks of this mechanism in CARINA are: ControlActivation and StrategySelection. A plan to solve the ReasoningFailure is built based on the new Goal. The plan is called FailureSolutionPlan. A FailureSolutionPlan can activate the metacognitive control. The ControlActivation task starts PlanExecution. StrategySelection is one of the tasks that comprise the plan. The StrategySelectiontask reads profiles of cognitive tasks and uses MetacognitiveStrategy to recommend computational strategies. The computational strategies are recommended to the CognitiveTask in order to solve the ReasoningFailure.
III. COGNITIVE MODELING LANGUAGE FOR CARINA
Computational cognitive modeling provides detailed descriptions of cognitive mechanisms [26] and embodies descriptions of cognition in algorithms and programs, based on the science and technology of computing [29], [30]. CARINA provides a programming language to create a cognitive model and an interpreter to execute the model. The programming language is called M++ and the development tool with the interpreter called MetaThink.
M++ is a domain-specific visual language specifically developed for modeling metacognition in intelligent systems. In M++ the specifications of the object-level and meta-level are supported in a metamodel configured according to the standard Meta-Object Facility (MOF) of Model-Driven Architecture (MDA) methodology. The architecture also constrains model building because it is based on MISM metamodel. This ensures that the resulting models are products of the theory behind the architecture.
Fig. 3. Screenshot of MetaThink –GUI
The concrete syntax was defined in order to make the language more usable. The icons were designed bearing in mind their usability when applied by users. Fig. 8 includes a representative sample of the elements of the M++ notation.
Fig. 4. Main elements in M++ notation.
In Fig. 4, section (A) shows the icons used to represent object-level tasks and section (B) displays icons representing elements that interact with the tasks at object-level. Section (C) contains the notation related to the tasks of introspective monitoring act from the meta-level on the object level. Section (D) displays icons representing elements that interact with the monitoring tasks. Section (E) displays icons representing the metacognitive control tasks and section (F) displays icons representing elements that interact with the tasks of
metacognitive control. In summary, M++ has approximately 20 notation elements for modeling metacognitive systems.
IV. CARINA COMPUTATIONAL IMPLEMENTATION
The CARINA architecture is an instance of MISM metamodel [22]. MISM metamodel is based on Meta-Object Facility (MOF) standard.
The MOF standard provides a framework based on metamodels to enable the development and interoperability of models to automate some stages of the development process of software applications. As can be observed in Figure 5.1, the architecture of CARINA’s framework is organized into four levels according to the MOF standard.
• Meta-MetaModel Level (M3). This levelcomprises meta-metamodel (MOF 2.0) that isused for implementation of the metamodel MISM(M2).
• Metamodel Level (M2). The metamodel MISM ispositioned at the M2-level in the MOFCARINA´s framework. Therefore, a Model that ispositioned at the M1-level can be modeled by themetamodel. MISM metamodel is specified usingMOF standard and implemented in EclipseModeling Framework (EMF).
• Model Level (M1). This level contains theconceptual models of CARINA that isimplemented according to the metamodelspecified at M2 level. In this sense, CARINA(M1 level) is a Metacognitive Arcitecture formonitoring and control of reasoning failures inArtificial Cognitive Agents.In the MOF metamodeling framework, thederivation of a model from its metamodel iscalled a ‘conformance.’ Through the conformanceprocess, a realization of concept in the MISMMetamodel in a new instance (object) in the
CARINA architecture at the M1 level can be achieved.
• User Model Level (M0). The User Model at theM0-level is the target model that is the aim of theMISM Metamodel. The derived target modelrepresents an Artificial Cognitive Agent in thereal-world. In MOF, the domain concept used in ametamodel is presented as a Class. The data for aClass is presented as an Object. As such, the datafor the Object are in turn presented as an Instancein User Model. End-Users manipulate real datausing Artificial Cognitive Agent generated by amodeling framework from M1, i.e. users cancreate and use models of entities from real world(M0), using the CARINA architecture (M1).
A. Description of the implementation of CARINA at Model
Level (M1)
CARINA is a metacognitive architecture developed under a paradigm oriented around cognitive cycles. Currently CARINA has the following implementations: Prolog,JavaScript and Java. The kernel of CARINA is named CARINA CORE. CARINA CORE is composed of a series of modules that define the basic elements of the computational implementation. The modules are organized into software packages. The CARINA modules are: memory, metacore.
1) Memory Package
The memory module contains the definitions of memory systems that can be used during the cognitive cycle. Memories involve the acquisition, categorization, classification and storage of information. The purpose of memory is to provide the ability to recall information and knowledge as well as events [31]. Memory System in CARINA is structured according to [32]–[34] as follows: Sensory Memory, Working Memory and Long Term Memory, see Fig. 5.
Fig. 5. Memory process in CARINA is based on [32]–[34]
Sensory memory is capable of retaining information for a very short period of time [34]–[37], retaining impressions of sensory information after stimuli have ended [36], [37]. Selective attention allows encoding of important information according to goals and expectations established in CARINA. Most of the information in sensory memory is not encoded in this way. The sensory memory is a temporary buffer which
briefly holds information not immediately attended [38], making it possible to attend to some of it slightly later.
Working memory is a system used for temporary storage of information during perception, reasoning, planning or other cognitive tasks [39]–[41]. A selective attention subsystem and a set of “Basic Cognitive Processing Units” (BCPU) compose CARINA’s Working Memory. In CARINA, a BCPU acts as a buffer among the cognitive processes that intervens in a cognitive loop.
Long-term memory stores information over the time [42]. In Carina, Long-term memory encodes information semantically for storage. Declarative memory (“knowing what”) is a memory of facts and events [43], and refers to those memories that can be consciously recalled (or "declared") [33]. In CARINA, declarative memory is subdivided into episodic memory and semantic memory.
• Episodic memory: In Carina episodic memory isimplemented as a Case-Based Reasoning System(CBR) and contains cases that represent events. Anevent consists of detailed sensory-perceptualinformation on recent experience [44], which includesperception, motor commands, and internal datastructures [45], [46]. Episodic memory is designed tosupport semantic memory [47], [48]. In this sense,episodic memory activity in Carina is analogous to thememory activity concentrated in the hippocampus.
• Semantic memory includes all acquired knowledgeabout the world [49], not contextualized in time andspace. In Carina, semantic memory has the followingproperties: (i) it simulates the activation of the frontaland temporal cortexes; and (ii) it consists of a “mentalthesaurus” implemented as an ontology. In this sense,the implementation of semantic memory in CARINAcorresponds to Tulving and Binder approaches [49],[50].
Procedural Memory is the memory of actions and behaviors of a system [51], including sequences, categories, rules and routes [52]. It is a non-declarative memory with nonconscious learning, which refers to a “how to” kind of information. In CARINA, this memory consists of a hybrid model based on a rule-based system to support sequences and provide a record of possible motor and behavioral skills.
2) Metacore package
The metacore module contains the basic elements and functions of CARINA, which are essential for the development of the meta-cognitive and cognitive processes of the system. Fig. 6 shows the basic package structure in CARINA.
The carina.metacore package facilitates the reuse of elements among different metacognitive mechanisms. carina.metacore facilitates the integration of metacognitive components in the design of intelligent systems because it is based on independent packages that share common design elements in metacore.
Fig. 6. Package model
The carina.objectlevel package contains the model an AI agent uses for reasoning about the world (i.e. agent’s environment) to solve problems [22].
The carina.metalevel package includes the components, knowledge and mechanisms necessary for self-monitoring and control of the learning and reasoning processes of an artificial intelligent agent. This package is composed of three sub-packages:carina.metalevel.core, carina.metalevel.monitoring and carina.metalevel.control.
The main objective of carina.metalevel.core is to simplify the complexity of carina.metalevel package. This package combines the classes that are common to the subpackages: carina.metalevel.monitoring and carina.metalevel.control. The carina.metalevel.monitoring sub-package includes mechanisms for detecting reasoning failures at the object-level. The carina.metalevel.control sub-package recommends to the object-level the best computational strategy to resolve a reasoning failure.
3) Validation
As with any scientific theory or engineered artifact, cognitive architectures require evaluation. However, because architectural research occurs at the systems level, it poses more challenges than does the evaluation of component knowledge structures and methods. In this section, we consider some dimensions along which one can evaluate cognitive architectures. In general, these involve matters of degree, which suggests the use of quantitative measures rather than all-or-none tests. Langley and Messina [53] discuss additional issues that arise in the evaluation of integrated intelligent systems.
4) Generality
Recall that cognitive architectures are intended to support general intelligent behavior. Thus, generality is a key dimension along which to evaluate a candidate framework. We can measure an architecture’s generality by using it to construct intelligent systems that are designed for a diverse set of tasks and environments, then testing its behavior in those domains.
The more environments in which the architecture supports intelligent behavior, and the broader the range of those environments, the greater its generality
5) Versatility
We can define the versatility of a cognitive architecture in terms of the difficulty encountered in constructing intelligent systems across a given set of tasks and environments. The less effort it takes to get an architecture to produce intelligent behavior in those environments, the greater its versatility.
6) Taskability
Generality and versatility are related to a third notion, the taskability of an architecture, which acknowledges that long-term knowledge is not the only determinant of an agent’s behavior in a domain. Briefly, this concerns an architecture’s ability to carry out different tasks in response to goals or other external commands from a human or from some other agent. The more tasks an architecture can perform in response to such commands, and the greater their diversity, the greater its taskability. This in turn can influence generality and versatility, since it can let the framework cover a wider range of tasks with less effort on the developer’s part.
V. COGNITIVE TASK AND MODEL
The real-world task that has been modeled in CARINA is the generation of an instructional plan from the instructional design domain. In this task, an instructional designer wants to design an instructional plan for a specific lesson of a course. The instructional plan will be applied individually and according to the learning style of a new student enrolled in such course. For this, the instructional designer searches among the plans that have been successfully applied to students with learning profiles similar to the new student. If the instructional designer finds an instructional plan then it is applied to the student, otherwise it will design a new instructional plan according to the characteristics of student. In summary, the task of the instructional designer is to determine the learning style of the student and select or build most appropriate instructional plan according to student’s learning style.
A. CARINA-based cognitive agents at User Model Level (M0)
Advancing in the modeling of the instructional planning
task, consider an instructional planner agent of a tutor module in an Intelligent Tutoring System. The instructional planner agent generates instructional plans, given the current state of the student, and monitoring the execution of the content plan in order to determine when to re-plan, or generate a new plan. In the context described, we have developed an Intelligent Tutoring System (ITS) called FUNPRO (FUNdamentos de
PROgramación) based on the MODESEC methodology [54], in order to show how the computational CARINA architecture can integrate cognitive and metacognitive functions.
The ITS is used as a case study on monitoring a plan´s execution. Monitoring plan execution is a high-level cognitive function. Monitoring and control of reasoning processes are two metacognitive functions supported in CARINA. A
metacognitive model of the instructional planner is constructed using the principles described in Section 2. The model of the object-level of the instructional planner is based on an extended Belief, Desire and Intention (BDI) framework [55]–[57].
In FUNPRO, the instructional planning task includes a subtask for recommendation of learning resources identified as
. The implementation of FUNPRO in CARINA includes defining the metacognitive mechanisms for introspective monitoring and meta-level control.
FUNPRO makes replanning to the instructional plan in the following cases: (i) if a resource for some reason cannot be deployed in the lesson, e. g. resource has the URL broken; (ii) if a recommended resource has received a poor evaluation; (iii) if the student obtains a low performance in the lesson. The instructional planning constantly refines the pedagogical strategy for each student using three types of recommendation strategies: (i) matching simple query, the search query in a simple SQL type; (ii) exclusive search is similar to (i), but excludes some results and; (iii) vote-based search, this strategy is based on the nearest neighbor algorithm. Fig. 7 shows the partial view of the metacognitive model for the instructional planning function in the FUNPRO, the view is focused in subtask.
Fig. 7. Metacognitive model generated in MetaThink using
M++ notation
The partial mapping from MOF metamodel (MISM) from CARINA to FUNPRO metacognitive model in M++ is listed as follows (Table I).
TABLE I. ITS – mapping table. CARINA - Concept Artifact of FUNPRO
ReasoningTask playResource
Strategy SQL_simple_query;
SQL_exclusion_query;
KNN_vote_based
Goal display_resource
ComputationalData play_resource_data
ReasoningTrace play_resource_trace
Profile playResourceProfile
Sensor playResourceSensor
ReasoningFailure UnavailableResource_
Failure;
URLbroken_Failure
FailureDetection anomalyDetectionInPl
ayResource
ProfileGeneration generatePlayResource
Profile
The meta-level intervenes in the process of instructional planning by deciding whether to continue reasoning for a better plan or execute the current plan. When a plan is generated, the meta-level analyzes the possibility of refining the plan (reasoning about the planning process) using as evidence the effectiveness of similar plans in the past. If the meta-level finds that expected performance of the current plan is sufficient to achieve the planned goals, then it proceeds to execute the plan. But if there are possibilities to improve the plan in a reasonable time, then the meta-level decides to continue planning further.
1) Implementation Issues
The instructonal planer in FUNPRO was implementedusing SWI-Prolog according to metamodel presented in Fig. 7. FUNPRO is an ITS for teaching Introduction to Programming
in Engineering. Fig. 8 shows a screenshot of the graphical user interface of FUNPRO.
Fig. 8. FunPro Screenshot
Metareasoning points (MRPs) define the reasoning tasks at object-level that are monitored and controlled using metacognition in meta-level. According to the object-level processes listed in Table II, instructional planning task is selected as a MRP. A failure in the execution of instructional
planning task could significantly affect performance of the
system. The meta-level keeps updated an abstract model of each MRP of object-level (e.g. self-model). The meta-level performs reasoning and decision making based on the self-model. Table II presents the main processes that are performed in FUNPRO for customizing instructional plans. Each process consists of several tasks. The tasks are implemented through rules. For better understanding, task names reflect what rules are specified in it.
TABLE II. Reasoning tasks in FUNPRO Process Reasoning task – Rule
Student profile identification
Verification of prerequisites and selection of lesson
Recommendation of pedagogical tactics
Recommendation and deployment of learning
resources
In Table I; S: Student - T: Pedagogic tactic - L: Lesson - R: Resource - U:
URL - I: Lesson component - Y: Learning style - C: Course - P: Performance
Introspective monitoring starts when playResourcegenerates new computational data (a resource path or an error message). The generation of a new computational data activates the sensor associated with the reasoning task. In the following code snippet we can see the implementation of the activation of a sensor in FUNPRO.
Reasoning failure detection starts when a sensor is activated. The meta-level gets updated observations from the sensor associated with the MRP, using the instructions
showActiveSensor(Action,Sensor) and update-SensorObservation (Sensor,Observation).Once the current reading of the active sensor is obtained, then meta-level checks whether the observation is consistent with the expectation of the sensor. In the following piece of code we can see the identification of a reasoning failure.
After reasoning failure is detected, the metalevel generates an explanation. There are three possible explanations for a reasoning failure occurred in
• Resource available refers to an unavailable resource whendeploying on FUNPRO.
• Inappropriate resource refers to a resource that was notadequate to the student profile.
• URL broke refers to an available or valid resource that has abroken URL.
Explanations can be generated in two ways: search for
known explanations and reasoning trace analysis.
Strategy search for known explanations queries for explanations given to reasoning failures in the past and then evaluates and prioritizes explanations, see the following piece of code.
The reasoning trace analysis strategy performs a more complex reasoning because it makes queries on the trace of the structures of reasoning performed by the failed task looking for anomalies, see the following piece of code.
Finally, the meta-level generates a goal based on the explanation in order to solve the reasoning failure.
Meta-level control starts after goal generation. The main function of metalevel control is to select the best available strategy to address the failure of reasoning. The selection of strategies receives as a parameter the Goal to be achieved and searches through the available strategies those which satisfy the Goal. Meta-level control then evaluates the performance of each strategy by selecting the best, in the following piece of code can be see the general implementation.
VI. CONCLUSION
CARINA is a metacognitive architecture for artificial intelligent agents. CARINA is derived from the MISM Metacognitive Metamodel. CARINA integrates self-regulation and metamemory with support for the metacognitive mechanisms of introspective monitoring and meta-level control. The main purpose of monitoring in CARINA is to provide enough information to make effective decisions in the meta-level control. The meta-level control in CARINA includes rules for the recommendation of the best strategy available to correct some reasoning failure at the object-level. CARINA provides a programming language to create metacognitive models and an interpreter to execute the models.
Currently CARINA has implementations in Prolog,JavaScript and Java. The kernel of CARINA is named CARINA CORE. CARINA CORE is composed of a series of modules that define the basic elements of the computational implementation.
The CARINA computational implementation allows the creation of intelligent software agents that conform to the CARINA metacognitive architecture and the MISM metamodel. This paper has demonstrated how a CARINA implementation can be added to an Intelligent Tutoring System (FUNPRO). A real-world task related to generation of an instructional plan from the instructional design domain was modeled in CARINA and implemented in the tutor module of FUNPRO. The added metacognitive capabilities allow the tutoring system to detect reasoning failure, to explain the failure and to select stratetegies for recovery.
The implementation of CARINA aims to simplify the process of developing a intelligent software agent and to permit the user to focus on the specifics of the application, hiding the complexities of the generic parts of the model.
ACKNOWLEDGMENT
This research was funded in part by the Internal Call for Sustainability of Research Groups of the University of Córdoba and by the project COL / 86815 – CONV. 647/15 of the Ministry of Information Technologies and Communications of Colombia (MinTIC).
REFERENCES
[1] Anderson, M., Oates, T., Chong, W., & Perlis, D. (2006). The metacognitive loop I: Enhancing reinforcement learning with metacognitive monitoring and control for improved perturbation tolerance. Journal of Experimental & Theoretical Artificial Intelligence,18(3), 387–411. http://doi.org/10.1080/09528130600926066
[2] Atkinson, R. C., & Shiffrin, R. M. (1968). Human Memory: A Proposed System and its Control Processes. Psychology of Learning and
Motivation - Advances in Research and Theory, 2(C), 89–195. http://doi.org/10.1016/S0079-7421(08)60422-3
[3] Baddeley, A. (1997). Human Memory: Theory and Practice. Book.Retrieved from http://books.google.com/books?hl=en&lr=&id=fMgm-2NXAXYC&oi=fnd&pg=PP7&dq=Human+memory:+theory+and+practice&ots=jL1h2MlDNi&sig=8_93t5UewpB9mhOZZw918NFBS3E
[4] Baddeley, A. (2003). Working memory: Looking back and looking forward. Nature Reviews. Neuroscience, 4(10), 829–839. http://doi.org/10.1038/nrn1201
[5] Binder, J. R., & Desai, R. H. (2011). The neurobiology of semantic memory. Trends in Cognitive Sciences.http://doi.org/10.1016/j.tics.2011.10.001
[6] Borst, J. P., & Anderson, J. R. (2015). Using the ACT-R cognitive
architecture in combination with fMRI data. (B. Forstmann & E.-J. Wagenmakers, Eds.)An Introduction to Model-Based Cognitive
Neuroscience. http://doi.org/10.1007/978-1-4939-2236-9_17 [7] Burgess, N., Becker, S., King, J. A., & O’Keefe, J. (2001). Memory for
events and their spatial context: models and experiments. Philosophical
Transactions of the Royal Society of London. Series B, Biological
Sciences, 356(1413), 1493–503. http://doi.org/10.1098/rstb.2001.0948 [8] Caro, M., Josyula, D., Cox, M., & Jiménez, J. (2014). Design and
validation of a metamodel for metacognition support in artificial intelligent systems. Biologically Inspired Cognitive Architectures, 9,82–104. http://doi.org/10.1016/j.bica.2014.07.002
[9] Caro, M., Toscazo, R., Hernández, F., & David, M. (2009). Diseño de software educativo basado en competencias. Ciencia E Ingenieria Neogranadina. jun2009, Vol. 19 Issue 1, p71-98. 28p. 4 Black and White Photographs, 10 Diagrams, 12 Charts., 19(1), 71–98. Retrieved fromhttp://search.ebscohost.com/login.aspx?direct=true&db=fua&AN=46984112&lang=es&site=ehost-live
[10] Cohen, N. J., Poldrack, R. a., & Eichenbaum, H. (1997). Memory for Items and Memory for Relations in the Procedural/Declarative Memory Framework. Memory, 5(1–2), 131–178. http://doi.org/10.1080/741941149
[11] Conitzer, V., & Sandholm, T. (2003). Definition and Complexity of Some Basic Metareasoning Problems. In arXiv preprint.
[12] Conway, M. a. (2001). Sensory-perceptual episodic memory and its context: autobiographical memory. Philosophical Transactions of the
Royal Society of London. Series B, Biological Sciences, 356(1413),1375–1384. http://doi.org/10.1098/rstb.2001.0940
[13] Cowan, N. (2010). Sensory and Immediate Memory. In Encyclopedia of
Consciousness (pp. 327–339). http://doi.org/10.1016/B978-012373873-8.00048-7
[14] Cox, M. (1997). An Explicit Representation of Reasoning Failures. In S. B. Heidelberg (Ed.), Case-Based Reasoning Research and Development
(pp. 211–222). [15] Cox, M. (2005). Metacognition in computation: A selected research
review. Artificial Intelligence, 169(2), 104–141. http://doi.org/10.1016/j.artint.2005.10.009
[16] Cox, M. (2007). Perpetual Self-Aware Cognitive Agents. AI Magazine,(2002), 32–51. Retrieved from http://www.aaai.org/ojs/index.php/aimagazine/article/view/2027/1920
[17] Cox, M., & Raja, A. (2012). Metareasoning: An introduction. In Cox and Raja (Ed.), In Metareasoning: Thinking about thinking (pp. 1–23). Cambridge. MA: MIT.
[18] Crowder, J. a, & Friess, S. (2010). Metacognition and Metamemory Concepts for AI Systems. In International Conference on Artificial
Intelligence, ICAI 2010. Las Vegas. [19] Crystal, J. D. (2009). Elements of episodic-like memory in animal
models. Behavioural Processes.http://doi.org/10.1016/j.beproc.2008.09.009
[20] D’Esposito, M., Ballard, D., Zarahn, E., & Aguirre, G. K. (2000). The role of prefrontal cortex in sensory memory and motor preparation: an event-related fMRI study [see comments]. Neuroimage., 11(5 Pt 1), 400–408.
[21] DeMarie, D., & Ferron, J. (2003). Capacity, strategies, and metamemory: Tests of a three-factor model of memory development. Journal of Experimental Child Psychology, 84(3), 167–193. http://doi.org/10.1016/S0022-0965(03)00004-3
[22] Eichenbaum, H. (2000). A cortical-hippocampal system for declarative memory. Nature Reviews. Neuroscience, 1(1), 41–50. http://doi.org/10.1038/35036213
[23] Fodor, J. A. (1975). The language of thought. New York (Vol. 4). http://doi.org/10.1016/0093-934X(77)90028-1
[24] Gais, S., & Born, J. (2004). Declarative memory consolidation: Mechanisms acting during human sleep. Learning & Memory, 11(6),679–685. http://doi.org/10.1101/lm.80504
[25] Horvitz, E. (1987). Reasoning About Beliefs and Actions Under Computational Resource Constraints. In Proceedings of the Third
Conference on Uncertainty in Artificial Intelligence (UAI1987).[26] Huet, N., & Mariné, C. (1997). Memory strategies and metamemory
knowledge under memory demands change in waiters learners. European Journal of Psychology of Education, XII(1), 23–35.
[27] Jeneson, A., & Squire, L. R. (2012). Working memory, long-term memory, and medial temporal lobe function. Learning & Memory (Cold
Spring Harbor, N.Y.), 19(1), 15–25. http://doi.org/10.1101/lm.024018.111
[28] Langley, P., Laird, J. E., & Rogers, S. (2009). Cognitive architectures: Research issues and challenges. Cognitive Systems Research, 10(2),141–160. http://doi.org/10.1016/j.cogsys.2006.07.004
[29] Machamer, P., Darden, L., & Craver, C. F. (2000). Thinking About Mechanisms. Philosophy of Science, 67(1), 1–25. http://doi.org/10.1086/392759
[30] Mikic-Fonte, F. A. (2010). A BDI-based intelligent tutoring module for the e-learning platform INES. In Frontiers in Education Conference
(FIE), 2010 IEEE (pp. 1–6). Washington, DC: IEEE Comput. Soc. http://doi.org/10.1109/FIE.2010.5673365
[31] Miller, E. K., Freedman, D. J., & Wallis, J. D. (2002). The prefrontal cortex: categories, concepts and cognition. Philosophical Transactions
of the Royal Society of London. Series B, Biological Sciences,357(1424), 1123–1136. http://doi.org/10.1098/rstb.2002.1099
[32] Miyake, A., & Shah, P. (1999). Models of working memory. Human
Performance Models for Computer Aided Engineering, 3, 203–214. http://doi.org/10.1017/CBO9781139174909
[33] Perlis, D., Cox, M., Maynord, M., Mcnany, E., Paisner, M., Shivashankar, V., … Caro, M. (2013). A Broad Vision for Intelligent Behavior: Perpetual Real-World Cognitive Agents. Annual Conference
on Advances in Cognitive Systems: Workshop on Metacognition in
Situated Agents, 2, 1–16. Retrieved from http://www.cs.umd.edu/~perlis/pubs/ACSvision21.pdf
[34] Piccinini, G. (2010). The mind as neural software? Understanding functionalism, computationalism, and computational functionalism. Philosophy and Phenomenological Research, 81(2), 269–311. http://doi.org/10.1111/j.1933-1592.2010.00356.x
[35] Raja, A., & Lesser, V. (2007). A framework for meta-level control in multi-agent systems. Autonomous Agents and Multi-Agent Systems,15(2), 147–196. http://doi.org/10.1007/s10458-006-9008-z
[36] Ramscar, M. (2010). Computing Machinery and Understanding. Cognitive Science, 34(6), 966–971. http://doi.org/10.1111/j.1551-6709.2010.01120.x
[37] Rao, A. S., & Georgeff, M. P. (1995). BDI agents: From theory to practice. In Proceedings of the first international conference on
multiagent systems ICMAS95 (pp. 312–319). Retrieved from http://scholar.google.com/scholar?hl=en&btnG=Search&q=intitle:BDI+Agents+:+From+Theory+to+Practice#0
[38] Scheutz, M. (2001). Computational versus causal complexity. Minds
and Machines, 11(4), 543–566. http://doi.org/10.1023/A:1011855915651
[39] Schmid, U., Ragni, M., Gonzalez, C., & Funke, J. (2011). The challenge of complexity for cognitive systems. Cognitive Systems Research, 12(3–4), 211–218. http://doi.org/10.1016/j.cogsys.2010.12.007
[40] Singh, P. (2005). EM-ONE : An Architecture for Reflective Commonsense Thinking. Dissertation of Doctor of Philosophy in Computer Science and Engineering - Massachusetts Institute of Technology.
[41] Sun, R. (2009). Theoretical status of computational cognitive modeling. Cognitive Systems Research, 10(2), 124–140. http://doi.org/10.1016/j.cogsys.2008.07.002
[42] Sun, R. (2012). The Motivational and Metacognitive Control in CLARION. In Integrated Models of Cognitive Systems.http://doi.org/10.1093/acprof:oso/9780195189193.003.0005
[43] Sun, R., Langley, P., Laird, J., & Rogers, S. (2009). Cognitive architectures : Research issues and challenges. Cognitive Systems
Research, 10(2), 141–160. http://doi.org/10.1016/j.cogsys.2006.07.004 [44] Sun, R., Zhang, X., & Mathews, R. (2006). Modeling meta-cognition in
a cognitive architecture. Cognitive Systems Research, 7(4), 327–338. http://doi.org/10.1016/j.cogsys.2005.09.001
[45] Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory (Vol. 1, pp. 381–403). New York, New York, USA: Academic Press. http://doi.org/10.1017/S0140525X00047257
[46] Tulving, E. (1983). Elements of Episodic Memory. Canadian
Psychology, 26(3), 351. http://doi.org/http://dx.doi.org/10.1017/S0140525X0004440X
[47] Turing, A. (1950). Computing machinery and intelligence. Mind,59(236), 433–460. http://doi.org/http://dx.doi.org/10.1007/978-1-4020-6710-5_3
[48] Unsworth, N. (2010). On the division of working memory and long-term memory and their relation to intelligence: A latent variable approach. Acta Psychologica, 134(1), 16–28. http://doi.org/10.1016/j.actpsy.2009.11.010
[49] VanPattern, B., & Williams, J. (2015). Theories in Second Language
Acquisition: An Introduction. Language Teaching.http://doi.org/10.2167/le128.0
View publication statsView publication stats