On Introspection, Metacognitive Control and Augmented Data Mining Live Cycles

On Introspection, Metacognitive Control and Augmented Data Mining Live Cycles

Abstract

We discuss metacognitive modelling as an enhancement to cognitive modelling and computing. Metacognitive control mechanisms should enable AI systems to self-reflect, reason about their actions, and to adapt to new situations. In this respect, we propose implementation details of a knowledge taxonomy and an augmented data mining life cycle which supports a live integration of obtained models.

Keywords: Metacognitive Modelling, Data Mining

1 Introduction

Cognitive computing is the development of computer techniques to emulate human perception, intelligence, and problem solving. Cognitive models are equipped with artificial sensors and actuators which are integrated and embedded into physical systems or ambient intelligence environments to act in the physical world. The goal is to have cognitive capabilities and to perform cognitive control (e.g., see [1]). To overcome problems in shared control (of, e.g., navigating robots [2]), direct communication (in natural language dialogue) between a human participant and a technical control architecture can be employed. This could be used for mutual disambiguation of multiple sensory modalities in a learning environment. As one of the major topics of sensory-based control mechanisms, automatic perception learning by introspection and relevance feedback could help in this disambiguation task. In order to pursue the idea of cognitive systems able to self-reflect, reason about their actions, and to adapt to new situations, metacognitive strategies can be employed.

In this paper, we will present the core idea of a metacognitive control model of machine learning with respect to problem solving capabilities to be exemplified by improving autonomous reaction behaviour.

We start by clarifying the term metacognition. Metacognition is cognition about cognition. It can, in principle, enable artificial intelligence systems to monitor and control themselves, choose goals, assess progress, and adopt new strategies for achieving goals.2 [4] associates metacognitive components with the ability of a subject (or an intelligent agent in general) to orchestrate and monitor knowledge of the problem solving process; [5] argues that metacognitive abilities correlate with standard measures of intelligence; [6] talks about systems that know what they are doing.

Here, we adopt the growing interest in metacognitive strategies3 for AI systems to build a metacognitive model for adaptable AI systems, which involves computational models of self-representation and self-awareness. Ontologies represent the knowledge groundwork for the self-representation of a system information state to be included into a metacognitive model.4 For example, McCarthy defines the term introspection as a machine having a belief about its own mental state rather than a belief about propositions concerning the world.

According to this explanation of metacognition we hypothesise that researchers in adaptable AI systems should investigate in metacognition because it can help us:

  1. address the difficulty to write down control management rules. Rules may not be obvious, tangible, or identifiable, or they may present an engineering overhead.

  2. provide self-improvement through adaptation and customisation.

  3. offer designs for never-ending learning.

  4. integrate a variety of previously isolated findings: dialogue architectures, finite state strategies, information states, (un)supervised learning, stacked generalisation, reinforcement learning, interactive learning, and embedded data mining.

Apart from its complexity, metacognition highlights an empirically tractable model creation and verification process.

2 Model, Introspective View and Control

We use the term model in the sense given by [7]:

To an observer , an object is a model of an object to the extent that can use to answer questions that interest him about .

can be the world or a specific sub-domain such as the football domain. To answer questions about the football domain, an has to be constructed. corresponds to an ontological knowledge base which contains facts about the sub-domain and the knowledge how to communicate the facts. This level of knowledge representation is basically implemented by state-of-the-art semantic technologies. Intelligent interaction systems for dialogical interaction with the Semantic Web (e.g., SmartWeb [15]) can be built on top of this representation of domain knowledge (e.g., dialogue and football knowledge).

Contemporary AI introduces the notion of ontologies as a knowledge representation mechanism (e.g., see [10]) for the operational AI models we are interested in. The object level represents the world and the domain of interest; in addition, the domain ontologies should contain mental concepts about communication and control structures; and for processing user feedback, a representation of natural communication (natural language dialogue) is required. When these concepts can be used to maintain an information state, a model of introspection can be derived from it. Then, self-reflective knowledge can be provided by the introspective AI system management facility which holds an introspective view of the object level. More precisely, an introspective view is obtained from introspective reports, i.e., interpretations of data records of process data as a description of the internal processes under observation. In this respect, we recognise introspection in the same way as done by [11]:

We view introspective reports as data to be explained, in contrast to the Structuralists’ view of introspective reports as descriptions of internal processes; i.e., we regard introspection not as a conduit to the mind but rather as a source of data to be accounted for by postulated internal processes.5

Thus, the introspective view can be implemented by the output of a meta-level data generalisation process while reporting on the object-level behaviour. Metadata providers decide which kind of information is to be included in the introspective reports. On the meta-level, meta-models can be generated with the help of machine learning and data mining algorithms. A knowledge taxonomy helps differentiate between the different knowledge levels, especially the knowledge levels obtained from the machine learning experiments.

2.1 Meta Knowledge Taxonomy

In order to integrate learning schemes—i.e. to learn meta-level action strategies from experience—we propose a meta knowledge taxonomy (figure 1). Consider a world () and a modeller () who exists in the world, and who can be a human or an intelligent computer agent. A knowledge taxonomy can be constructed to include the modelling of the world and the modeller (according to some articles in [12]). In this paper, we provide the implementations of this knowledge taxonomy by using semantic technologies and machine learning.

Figure 1: Meta Knowledge Taxonomy for metacognition

The model of the world () is used to answer questions about . In order to answer questions about the modeller himself, we introduce a model of the modeller (). is the self (reflective) knowledge of the agent in the world. Before explaining the crucial layers of introspective knowledge for complex AI systems, we should explain our idea of how to map and implement the first four layers. The world is the application we have in mind, with the capability to adapt to new environmental conditions. Thereby, the processing system is the modeller . Accordingly, is the processing system’s knowledge basis, the ontology terminology box of the AI system’s application domain. is the internal state of the dialogue system which we implement as information state, consisting of assertion box instances, according to the ontology. It represents the self knowledge of the system in the running state. If the information state contains information about the system itself, the modeller’s self knowledge can be called reflective.

is the model of the world knowledge; it contains the meta knowledge in order to reason about the questions concerning the world knowledge. Some typical questions for ontological knowledge bases are whether the classes and relations adequately describe the application domain, and whether the descriptive representation of domain processes provides a motivated conclusive representation of the situation in terms of content-describing features. We implement this meta knowledge layer with machine learning models: if the models created by the attributes derived from ontology instances have positive evaluation characteristics (for example, high cross-validated classification accuracy or reasonable symbolic association rules or decision trees), we adequately describe the world knowledge by meta knowledge. (A task-based evaluation of ontologies in a specific application domain is meta knowledge, too.) is the knowledge that can be extracted from the processing system while running the system in the current environment. This self-reflective knowledge can be used to adapt to other processing strategies, for example control signals, if the current one fails. At this layer, we are able to recognise how all other knowledge layers work together to performing a particular task in the AI system’s application domain.

Both and can be used to build decision-oriented operationalised management rules. Decision-oriented means that any of the reaction duties are directly triggered or effected. Operationalised rules means that the control rules derived from the machine learning models are in a directly executable format (e.g., association rules) or can be translated into these. The operationalisation itself can be undertaken manually or automatically. To sum up the intention of the meta knowledge taxonomy for metacognition. The modelling by a knowledge taxonomy provides abstract solution for the problem of how

  • to monitor system performance;

  • to adapt a problem solving strategy according to performance classification;

  • to build operational machine learning models.

3 Augmented Data Mining Live Cycle

The implementations of the knowledge taxonomy are given by the processing system, the (ontological) knowledge basis, the information state, the meta knowledge by ML models, and the introspective ML models. Thereby, the theory combines top-down approaches (i.e., ontological knowledge representation) with bottom-up approaches (i.e., empirical process data model exploitation). The later means information state features aggregation and data mining by combining declarative and procedural knowledge. Metacognitive control is the application of the introspective knowledge gained on the meta-level by controlling the object-level, as illustrated in figure 2. According to control theory, we are not only able to vary parameters of the object level control in real-time, but augment the object-level (cognitive) reasoning process by learned meta-models. Hence, the metacognitive control idea includes planning, monitoring, authoring, integration, and evaluation.

The last two steps, integration and evaluation, are implemented by augmenting the data mining life cycle to support a live integration of obtained models. We call this additional step the (automatic) operationalisation of learned meta models. Figure 3 illustrates the Cross Industry Standard Process for Data Mining cycle6 and includes our augmentation. In the modelling phase, various modelling techniques are selected and applied. The modelling phase is finished when one or more models, which appear to be of high quality at least from a data analysis perspective, have been built. These models then need to be evaluated before their deployment. In the evaluation phase we use the models to review the model building process. This evaluation is done by running the system on unseen supervised data or by reinforcement learning experiments. Finally, at the end of the evaluation stage, a decision has to be reached as to whether to use the data mining results obtained. Then a new model is deployed and used in the domain or business units.

Figure 2: The meta-level control is established by the embedding of introspective knowledge for control. The augmented data mining cycle is shown in figure 3.

The CRISP cycle closes with the evaluation of the deployed system in the real application context (domain/business understanding), whether it performs well, or not. In fact, this is a kind of metacognitive process conducted by the domain experts. The introspective mechanism represents a new phase between evaluation and (human) domain/business understanding. It automatically optimises the behaviour of the deployed system and provides hints for human understanding by generating transparent metamodels of the system’s performance, for example, introspective association rules and decision trees. The cycle now includes the additional step (automatic) operationalisation before it closes.

Our aim to integrate the introspective mechanism in order to extend the data mining cycle by a new phase where system introspection is integrated, resulted in a new step of the data mining life cycle, i.e., (automatic) operationalisation. The introspective models are directly used in conjunction with the former decision making models for action taking. As a result, the augmentation of the CRISP cycle represents a tractable metacognitive model creation and verification process. In subsequent applications of the augmented CRISP cycles, the introspective models can be combined with the models of the former CRISP process.

Figure 3: Adapted CRISP data mining cycle. CRISP is characterised by its independence from the application domain and the algorithms used. This makes it suitable as base data mining cycle where metacognitive aspects (also independent from domain and algorithms) are included.

It is important to note that empirical machine learning models are pattern patching systems; we expect the behaviour to be improved by drawing an analogy to a past experience which materialises as patterns to be mined. These patterns do not necessarily follow logical rules in terms of a higher order logic—but instead, they should follow at least the causal implications of a propositional logic which helps to implement reactivity based on learned causality. All patterns to be mined can be regarded as introspective reports on the application or business domain.

4 Conclusion

The question we investigated was about the scope and usefulness of a metacognitive model. In order to develop a computational introspective model, empirical machine learning models can be investigated. This should augment cognitive capabilities of adaptable AI systems, especially in the reasoning phase before action taking, which we believe requires to a great extent metacognitive instead of cognitive capabilities.

Similar methodology in computation has received great attention for uncertainty handling, control in decentralised systems, scheduling for planning in real-time, and meta-level reasoning in general [13]. Applications are to be found in the contexts of large-scale natural language processing architectures for texts (e.g., UIMA [14]), and dialogical interactions with the Semantic Web (e.g., SmartWeb [15] integrating extensive ontological groundwork [16] for self-representation of an information state to be included into a metacognitive model). The metacognitive control and augmented Data Mining Cycle proposed here will be integrated into a new situation-aware dialogue shell for the Semantic Access to Media and Services in the near future—to handle, fore and foremost, the access to dynamic, heterogeneous information structures.

Acknowledgements.

This research has been supported in part by the THESEUS Program in the Core Technology Cluster WP4 Situation Aware Dialogue Shell for the Semantic Access to Media and Services, which is funded by the German Federal Ministry of Economics and Technology under the grant number 01MQ07016. The responsibility for this publication lies with the author.

Footnotes

  1. email: sonntag@dfki.de
  2. For example, students preparing for an exam judge about the relative difficulty of the learning material and use this for study strategies. The resulting reasoning task is a second-order reasoning process about the own learning abilities called meta-reasoning or, more generally, metacognition.
  3. IBM Autonomic Computing Initiative, http://www.research.ibm.com/autonomic/,
    and, e.g., DARPA Information Processing Technology Office on Cognitive Systems,
    http://www.darpa.mil/ipto/thrust_areas/thrust_cs.asp.
  4. [8] outlines that for intelligent behaviour, a declarative knowledge model must be created first. Examination of, e.g., own beliefs would then be possible when the beliefs are explicitly represented. McCarthy sees introspection as essential for human level intelligence (and not a mere epiphenomenon) [9].
  5. This important quote basically states that metacognition as proposed here is not a reconstruction of the respective human intelligence apparatus—in accord with technical cognitive AI system research.
  6. See http://www.crisp-dm.org.

References

  1. Beetz, M., Buss, M., Wollherr, D.: Cognitive technical systems — what is the role of artificial intelligence? In Hertzberg, J., Beetz, M., Englert, R., eds.: Proceedings of the 30th German Conference on Artificial Intelligence (KI-2007). (2007) 19–42 Invited paper.
  2. Ross, R.J., Shi, H., Vierhuff, T., Krieg-Brückner, B., Bateman, J.A.: Towards dialogue based shared control of navigating robots. In Freksa, C., Knauff, M., Krieg-Brückner, B., Nebel, B., Barkowsky, T., eds.: Spatial Cognition. Volume 3343 of Lecture Notes in Computer Science., Springer (2004) 478–499
  3. Pantic, M., Pentland, A., Nijholt, A., Huang, T.: Human computing and machine understanding of human behavior: a survey. In: ICMI ’06: Proceedings of the 8th International Conference on Multimodal Interfaces, New York, NY, USA, ACM Press (2006) 239–248
  4. Derry, S.: Strategy and expertise in solving word problems. In: Cognitive strategy research: From basic research to educational applications. Springer-Verlag, NY (1989) 269–302
  5. Davidson, J., Deuser, R., Sternberg, R.: The role of metacognition in problem solving. In: Metacognition. The MIT Press. (1994) 207–226
  6. Brachman, R.J.: Systems That Know What They’re Doing. IEEE Intelligent Systems 17(6) (2002) 67–71
  7. Minsky, M.: Matter, mind, and models. M. L. Minsky (ed.) Semantic Information Processing (1968)
  8. McCarthy, J.: Programs with Common Sense. M. L. Minsky (ed.) Semantic Information Processing (1968)
  9. McCarthy, J.: Making robots conscious of their mental states. In: Working Notes of the AAAI Spring Symposium on Representing Mental States and Mechanisms, Menlo Park, CA (1995)
  10. Fensel, D., Hendler, J.A., Lieberman, H., Wahlster, W., eds.: Spinning the Semantic Web: Bringing the World Wide Web to Its Full Potential. In Fensel, D., Hendler, J.A., Lieberman, H., Wahlster, W., eds.: Spinning the Semantic Web, MIT Press (2003)
  11. Nelson, T.O., Narens, L.: Metamemory: A theoretical framework and new findings. In: G. H. Bower (Ed.) The Psychology of Learning and Motivation: Advances in Research and Theory. Volume 26. Academic Press (1990) 125–169
  12. Minsky, M.: The society of mind. Simon & Schuster, Inc., New York, NY, USA (1986)
  13. Anderson, M., Oates, T., eds.: 2005 AAAI Spring Symposium on Metacognition in Computation, March 21-23 2005, Stanford, California, USA. In Anderson, M., Oates, T., eds.: Metacognition in Computation. Volume SS-05-04 of AAAI Technical Report., AAAI Press (2005)
  14. Ferrucci, D., Lally, A.: UIMA: an architectural approach to unstructured information processing in the corporate research environment. Natural Language Engineering 10 (2004) 327–348
  15. Sonntag, D., Engel, R., Herzog, G., Pfalzgraf, A., Pfleger, N., Romanelli, M., Reithinger, N.: SmartWeb Handheld—Multimodal Interaction with Ontological Knowledge Bases and Semantic Web Services. In Huang, T.S., Nijholt, A., Pantic, M., Pentland, A., eds.: Artifical Intelligence for Human Computing. Volume 4451 of Lecture Notes in Computer Science., Springer (2007) 272–295
  16. Oberle, D., et al: DOLCE ergo SUMO: On foundational and domain models in the SmartWeb Integrated Ontology (SWIntO). Web Semant. 5(3) (2007) 156–174
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
214454
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description