الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 141 |  |  |
| | | from | 141 | | |
Abstract
his chapter presents a summary and conclusions to the proposed work highlighting interesting points in the thesis and finally suggest directions for future work.
Chapter 8: Summary, Conclusions and Future Works
8.1 Summary
For an autonomous intelligent agent, understanding a new current social situation (Target) needs to recall an appropriate past social experience relevant one (Base). Therefore, the intelligent agent must be able to determine the event(s) that can be performed in the current situation, without violating the social normsion for that situation. The challenges provided in learning social norms require the agent not only to have detailed knowledge about different social situations but also keep track of the temporal presidency of the events included in that social situations. This is one challenge many knowledge representations and learning techniques do not know how to deal with.
In this instance, this thesis proposed a novel computational model for human episodic learning as a cross-domain analogical learning that augments autonomous agents’ mental capabilities with a procedural way to transfer episodic knowledge from one domain (Base) to another particular domain (Target), on the basis of similarity. By extracting new experienced knowledge from pre-existing socio-cultural cognitive scripts and from a pre-established immense commonsense knowledge base. It requires that the two domains are structurally and contextually similar to each other. This knowledge is necessary for the AI agents to develop learning for the current faced social situation (Target).
8.2 Conclusions
The contributions of this thesis can be briefly concluded as follows:
8.2.1 Analogical Reasoning Methodology
Based on the comparative study of different analogical reasoning methodologies for learning systems conducted in Chapter 3, the cognitive script is proposed as the current knowledge representation where it describes how sociocultural schemas are expected to unfold. Thus, capturing social and cultural norms, and preserve the temporal progression and causal relations between the events in the schemas are what confer context to a specific cognitive script.
8.2.2 Information Retrieval Algorithm
Consequently, Chapter 5 discusses in detail the proposed model architecture. The proposed model depends mainly on two modules, the retrieval and learning modules. First, the retrieval module receives the current social situation in the cognitive script representation (Target) as an input and consults a knowledge base of cognitive scripts of different social situations. Then the most similar social situation script (Base) to Target script is returned using a context-based information retrieval algorithm (Pharaoh) for cognitive scripts.
In this thesis, an improvement of Pharaoh algorithm has been implemented based on a comparative study as shown in Chapter 7. That leads to achieve a better precision and recall (83.33%), and reduce the average time (to 1.06 seconds) required for the retrieval process.
8.2.3 Learning Algorithm
Afterwards, the learning module on the basis of analogy uses the retrieved Base script to enrich Target script with new events and relations that can be possibly learned. The key idea of learning by analogy is to transfer knowledge from one situation (Base) to another situation (Target) using mechanisms (crossover and mutation) inspired by evolutionary computation. Furthermore, for better expanding Target script with all possible events that can occur in it, the learning module gains an advantage of learning from a pre-existing commonsense knowledge base (ConceptNet).
For providing more detail about how the proposed model works. A case study on enriching “entering a stadium” social situation is presented in Chapter 6. As shown in Chapter 6, the retrieval module selected the Cinema script (Base) as the most contextual similar structure to the Stadium script (Target), with similarity value 0.31, in entering a place (stadium/cinema), buying a ticket, entering the seating area, sitting down, watching game/movie, and at the end leaving the place. After that, the learning module transferred seven new relations and two new events from the Cinema to the Stadium script by the crossover mechanism. Finally, 30 new events and 53 new relations enriched the Stadium script by crossover with all similar scripts to the Stadium script in the evolved script-base and by mutation with subsequent events from ConceptNet. Which, in turn, show that the proposed model successfully enriched the social dynamics associated with the “entering a stadium” situation.
8.2.4 The proposed Model Roles
Regarding above, the proposed model run in a creative way by adapting knowledge to the new situation and/or coming up with new knowledge as a result of blending two or more previous experiences.
One advantage of the proposed model is its ability to deal with defeasible knowledge, the knowledge that describes something that is not always true (e.g. “buying a popcorn is a subsequent event of entering a cinema event”).
Another merit of the proposed model is that its representation scheme is very expressive and straightforward. In addition, the representation grounded in structured English fragments and preserves the temporal progression of a specific context. It can be considered and rated as the first computational model that utilizes the idea of evolutionary mechanisms (crossover and mutation) in learning by cross-domain analogy from cognitive scripts. Furthermore, the learning algorithm is based on the inexact matching between events, and it can soundly infer the one-to-one mappings (isomorphism) between scripts’ parameters.
Finally, the proposed model has been implemented in Prolog and compiled as a library for easily deploying it to work on other applications. In addition to the main format for storage within the model is XML for facilitating the machine-readability. Accordingly, the proposed model can be used in many applications that exhibit human-like episodic learning capability such as:
1. building interactive agent or a cognitive robot that can naturally deal as humans in social situations or can exhibit many cognitive capabilities such as sensing, internal simulation, and prediction, learning action models and retrospective reasoning and learning
2. building computational narrative intelligence system for automatic story generation
3. building a cognitive tutor that can create simple real examples from real life to explain complex scientific things.
8.3 Future Works
The future works that are related to this thesis work can be briefly recommended as follows:
1. Building a large knowledge base of multi-branched cognitive scripts from pre-existing commonsense knowledge bases. This is to start learning from a large initial foundation.
2. Defining meta-rules that infer which subsequent event, out of a set of subsequent events to the current event, is more likely to occur than another in a particular situation.
3. Deploying the proposed model to build a complete cognitive architecture.
4. Developing an elaboration algorithm that can automatically evaluate the result of the learning module and remove unbelievable (odd) events and relations.
5. Defining explicitly causal connections between temporal ordered events.
6. Developing a recognition algorithm that can distinguish and differentiate between causal and temporal relations that connect events.