Predictive Agent Learning Driven By Temporal Data Mining Based On Information Decay

Artificial intelligence is currently one of the most interesting research fields in the computer science community. AI is intended both to help researchers understand how the brain makes decisions and to augment the decision-making process for organizations. On the heart of AI there are intelligent computational agents—simply called agents—that have the goal of building systems that can adapt to their environments and learn from their experience. This goal has attracted researchers from various disciplines and resulted in invention of variety of learning techniques that are transforming many industrial and scientific fields nowadays.These learning methods have proven to be of great practical value in a variety of application domains, especially in data mining, which is the application of AI learning techniques in extracting knowledge from large databases that contain valuable implicit regularities that can be discovered automatically. These methods, also known as machine learning techniques, are interpreted as the acquisition of structural descriptions—known as a model—from past data. The model can then be used as the knowledge representation of the agent for prediction, explanation, and understanding.Most predictive agent learning methods are based on the assumption that the historical data involved in building and verifying a model are the best estimator of what will happen in the future. However, the relevance of past to the future depends on the application domain in a specific timeframe. The reason of this static treatment of data is due to the fact that these learning methods abstracted the agent learning from human learning and left behind the forgetting factor that affects the acquired knowledge after learning. Therefore, it is imperative to integrate the forgetting in existing agent learning methods in order both to approach the human brain decision making and to improve learning performance of agents.It is in that spirit that in this thesis we present an abstraction of human forgetting that can be used by agents to improve their predictive learning capability and to assist them in deciding about the future in human like nature. We called that forgetting abstraction information decay learning technique.To reveal the fruition of the information decay learning approach, rather than specifying the theoretical concepts behind it only, we use data mining to experiment and provide practical findings of it. As there are a couple of data mining subfields in existence, we have preferred the use of temporal data mining that deals with the extraction of knowledge from temporal database due to its provision of time required for the computation of information decay.The thesis describes a multi-agent system architecture that is implemented for experimentation of the forgetting abstraction. The described architecture reveals how the agent and data mining technologies are integrated in a single system as a means of providing a single view of using information decay in discovering knowledge from large databases to enhance the learning capability of agents.To provide as sounding proof of this hypothetical theory, the information decay is implemented in the impurity measure of the ID3 decision tree learning algorithm, and it is tested on a dataset of 940 examples gathered from Oracle database. Two knowledge representation models of a predictive learning agent are developed:in one case ignoring the decaying nature of information in data, and in the other case considering the information decay. The Accuracy, Fl-measure, and Receiver Operating Characteristic curve are the metrics used for measuring the learning performance of the agent. The relevance of the concept is proven by comparing the learning performances of a predictive learner agent that uses the two models at different times to make prediction over an unseen subset of the collected dataset.