The Massive Memory Project

The Massive Memory Project

GrantReference: CICYT TIC90-0801-C02-01 Title: Project AMP: Development of a learning system based on a massive memory architecture Project Responsible: Enric Plaza i Cervera Institution: Consejo Superior de Investigaciones Científicas Institute: Centre d'Estudis Avançats de Blanes (former name of IIIA) Period: from 90.12.12 to: 93.12.12 Keywords: Artificial Inteligence, Machine Learning, Analogy, Case-based Reasoning, Representation Languages

Project Team
Project Goals
Project Results
Publications on the Web

This is a finished (1990-1993) project but a follow-up project is currently building upon it: The ANALOG Project.

Project Team

Josep Lluís Arcos (IIIA)
Eva Armengol (IIIA)
Ulises Cortés García (LSI, UPC)
Beatriz López Ibàñez (IIIA)
Ramon López de Mántaras (IIIA)
Enric Plaza i Cervera (IIIA)
Walter van de Velde (IIIA & AI-Lab, VUB)

Project Goals

There is a white paper on the Massive Memory Project. The goal of the Massive Memory Project (MMP) is developing an experimental framework for building systems using different problem-solving methods and different ways of learning from experience. Moreover, the specific goal is to show that the right level for integration of problem-solving methods, learning from experience is an architecture based on an active memory. All represented structures are memory structures: cases, generalizations, domain models, problem-solving methods, learning methods, instances of problems solved by one method, etc. We can summarize the Massive Memory Project in six working hypothesis. The research around MMA is directed by a series of hypothesis:

Learning is a type of meta-level inference. [Meta-level Inference Hypothesis]
Learning requires a self-model of the architecture [Self-model Hypothesis]
Self-models are knowledge-level descriptions. [Knowledge-level Hypothesis]
A specific learning method embodies a particular theory (knowledge) about how to improve the system's behavior. [Learning Meta-theory Hypothesis]
Learning methods can be described as reasoning tasks by means of knowledge-level models. [Learning Components Hypothesis]
Reification of knowledge-level descriptions allows the explicit representation of learning methods as meta-level inference methods. [Explicit Method Hypothesis]
Reflection offers a conceptual framework for describing and integrating learning and reasoning. [Reflection Hypothesis]
Reflection is a mechanism for integrating learning and reasoning in a computational architecture. [Reflective Integration Hypothesis]

Project Results

The main results of the project are theoretical and technical

Theoretical results

New learning methods

Induction of decision trees based on a new entropy measure. Attribute selection was guided by this new measure that estimates distance between the teacher's partition and the partition generated so far by the decision tree. This entropy measure was tried on datasets on hepatitis and lung cancer with an improvement on results compared to classic methods for decision tree induction. R. López de Mántaras (1991), A Distance-based Attribute Selection Measure for Decision Tree Induction. Machine Learning Journal. Vol. 6, n. 1, pp. 81-92.
BOLERO is a metalevel system for learning control knowledge. BOLERO uses case-based reasoning for learning strategic plans for guiding from the meta-level the diagnositc precedure of an expert system in pneumonia diagnosis. Furthermore, BOLERO learns to identify goal states (i.e. when to stop the diagnosis process). B. Lopez, E. Plaza (1993), Case-based planning for medical diagnosis. Lecture Notes in Artificial Intelligence 689, 96-105. Springer Verlag.

Integration of learning methods

Integration of learnming methods with problem solving systems is crucial for (a) supporting learning from past problem solving experiences and for (b) improving problem solving in future problems. Below there is information about a workshop and a position paper on the topic.

ECML-93 Workshop on Integrated Learning Architectures.
E. Plaza, A Aamodt, A. Ram, W. van de Velde, M. van Someren (1993), Integrated Learning Architectures. Lecture Notes in Artificial Intelligence n. 667. Springer-Verlag, pp. 429-441. (This is a position paper on the topic. Read it right here)
BOLERO is a meta-level system developed in the project that usesthe reflection mechanism involved in object-level/meta-level interaction to integrate learning and problem sovling. BOLERO consists in
- a rule-based object-level capable of approximate reasoning for diagnosis tasks, and
- a meta-level case-based planner that learns how to decide, at each instant, which object-level dignosis are worthwhile to be cosnidered.
Object-level contains domain knowledge for problem solving (how to deduce plausible diagnoses) from the patient data, while the meta-level contains (in fact, learns) strategic knowledge (planning from all possible goals which are likely to be useful). BOLERO's learning regards improving the system's focus in problem solving. [Case-based planning for medical diagnosis López & Plaza 93].
The case-based planner is used to control the search space of the object-level improving the whole system efficiency. The contro regime is reactive: every time there is a new information in the object-level the control is passed to the meta-level assuring that the system is dynamically capable of generating new strategic plans for the currently available information. As result, BOLERO is capable of a reactive behavior: any change in situation is able to provoke a cnage in the strategic plan and thus in the goals to be pursued by the system. BOLERO is the first planning system we are aware of that combines case-based and reactive capabilities, and this is achieved by the reflective organization of the systemn [Reactive Planning through the Integration of a Case-Based System and a Rule-Based Systetem López 93]. This is the result of the Ph.D. of Beatriz López Ibàñez: Learning strategic knowledge for diagnositc systems.
LINNEO system was an expeirment in integrating explanation based learning (EBL) and similarity-based learning (SBL). LINNEO constructs a classification hierarchy from a set of (possibly error-prone) observations and a domain (possibly incomplete) theory. M. Martin, R. Sangüesa, U. Cortés (1991), Knowledge acquisition combining analytical and empirical techniques. In L. A. Birnbaum, G.C. Collins (Eds.), Proc. Eight International Workshop on Machine Learning, p. 657-661.

Development of a Memory Architecture

The Massive Memory Architecture (MMA) is a step forward from the BOLERO experiment: in addition to the integration of a learning method and a problem solving system, MMA it addresses the issue of different learning methods with problem solving. For achieving this, our hypothesis was that learning methods could be described at the knowledge level. E. Plaza, J.L. Arcos (1993a), A Reflective Architecture for Integrated Memory-based Learning and Reasoning. First European Workshop on Case-Based Reasoning EWCBR-93. (University of Kaiserslautern Germany) 1 November 1993. pp. 329-334. To be published in Lecture Notes in Artificial Intelligence, Springer Verlag.
We used knowledge modelling techniques, used by the knowlege acquisition community to develop expert systems, to analyze and describe learning methods. Analyses of explanation based learning systems [Armengol y Plaza 93b] and case-based reasoning systems [Armengol y Plaza 93a] performed show that knowledge modelling is a useful and powerful tool to analze, describe, and specify (existing or new) learning methods. Since problem solving has already been analyzed using knowledge modelling, this approach has the advantage of a uniform representation for both learning and problem solving.
Having uniformely represented both learning methods and reasoning methods it is possible to design the MMA that implements them in a uniform way (see Technical Results section below). An important result from the work in BOLERO and the knowledge modelling analysis is that learning is a type of meta-level inference (with respect to inference in the problem solving level). A system with learning integrated with problem solving requires an introspection capability (analyzing the results and behavior of the system itself) and self-modification (in order to obtain an improved ebhavior of the system itself). This reflective nature of learning processes has been essential to develop NOOS, a representation language for implementing systems that integrate reasoning and learning methods, explained in next section.

Technical results

Language

The main technical result is the specification and implementation of NOOS, an object centered representation language with reflective capabilities. NOOS allows to uniformely represent reasoning and learning methods. We have shown it is easy to program and represent in NOOS inheritance methods (in several variations) and derivational analogy and other case-based methods (with different case retrieval and selection methods and criteria). These methods are explicit and programmable, instead of being built-in and fixed (as is the usual approach, sometimes for each application domain) [Plaza 92c]. NOOS allows a clear separation and distinction between different types of knowledge, i. e. levels and meta-levels and their interaction are clearly established and the programmer can describe the place each piece of knowledge occupies and it's relation to the others. Different learning methods can be then described in NOOS (see below and E. Plaza, J.L. Arcos (1993b), Reflection and Analogy in Memory-based Learning. MSL-93 Second International Workshop on Multistrategy Learning. Harpers Ferry, USA.)

The second version of the NOOS language is currently being developed in the ANALOG project.

Architecture

The memory architecture implemented is integrated and flexible: (1) methods themselves are represented declaratively as memory structures (clichés), and (2) memory processes are reified in the NOOS language and allows their explicit modification by programming. The memory architecture consists of the NOOS language, the reflective constructs of which support access and manipulation of all language objects, plus a set of memory retrieval methods that support searching and collecting problems solved in the past in the form of memory structures and use them as objects of the NOOS language usable in solving new problems (achieving learning). Thus "system behavior" achieved by methods is not procedural and opaque but stored in memory as declarative memory structures allowing their retrieval and inspection, so that the system can reason about them and their results, and reuse them when appropriate in future situations. This notion of memory is that which distinguishes the NOOS from other languages and is what permits to integrate learning methods (as methods with specific inference strategies based on searching the memory of past experiences). [Plaza y Arcos 93c].

Methods

Regarding methods, we have developed reasoning and learning methods described as high level specifications (clichés) using knowledge-level analysis. Among this methods are: gaol-driven analogy, derivational analogy, analogy by determinations, case-based planning, variants generate-and-test methods (case-based and using induction to acquire the knowledge needed to generate and test hypotheses) [Plaza y Arcos 94c].

The BOLERO system was applied to the domain of pneumonia diagnosis.
CHROMA, an application for refining proteins using case-based and inductive methods was developed using NOOS. Currently a new version of CHROMA (being developed inside the ANALOG project) features a meta-level methods capable to decide which method is best in each situation.

Publications

Here is a complete list of publications and reports of the AMP project (35 items).

Some publications about this project are readable in the web:

Flexible Integration of Multiple Learning Methods into a Problem Solving Architecture (Enric Plaza, J L Arcos). This is an extended version of the paper in the Proceedings of the European Conference on Machine Learning in 1994.
Reflection, Memory and Learning (Enric Plaza, J L Arcos). This is an extended version of the paper in the Proceedings of the Workshop on Multistrategy Learning in 1993.
Integration of Learning into a Knowledge Modelling Framework (Enric Plaza, J L Arcos). This is the paper in the Proceedings of the European Workshop on Knowledge Acquisition in 1994.
Integrating Induction in a Case-based Reasoner, (Eva Armengol, Enric Plaza). This is the paper in the Proceedings of the European Workshop on Case-based Reasoning in 1994

For any question or information just ask

enric@iiia.csic.es