Understanding Reinforcement Learning Theory for Operations Research and Management

Introduction

Since the introduction of computers, humans have shifted laborious tasks which were once manually operated, unto machines for processing. Decades after the introduction, many intelligent approaches have been discovered and implemented. On daily basis, algorithms supporting such approaches are upgraded to suit modern trends of information processing. The inception of the internet for example, has introduced a higher degree of complexity in data (e.g. video, text, etc), prompting algorithm enhancements. Today, news production agencies such as BBC, CNN, indirectly build up the quest for algorithm upgrades by publish news items in the form of text, video, audio for general public consumption. With scientific estimate of institutional amorphous data close to eighty percent (80%), it has become prudent for organizations to consider amorphous data in business intelligence modeling. Although computers are able to extract, process and store huge amounts of data, human intelligence is by far, the highest form of intelligence. Basically, humans learn by extracting stimuli from immediate surroundings and using extracted stimuli to generate required actions. Prior to generating actions, inherent human instinct, processes extracted stimuli and applies prior knowledge whilst adjusting same (knowledge) in the learning process. Naturally, humans learn from tasks they perform and use acquired knowledge to perform same or similar task in the future. The learning process may be guided through a teacher, partly through a teacher or by the human being itself. In the context of machines, the human being is a learning agent. Learning agents in general, acquire input from their immediate surroundings (i.e. environment), process input with/without initial knowledge (initial weights) and generate/adjust knowledge (updated weights) with new information. A learning agent that performs a task all by itself is termed, an unsupervised learning agent. Similarly, a learning agent that performs a task partly by itself and partly through a teacher is termed, a semi-supervised learning agent. Learning agents that perform a task one hundred percent through a teacher is known as a supervised learning agent.

Amongst the learning paradigms, unsupervised learning is most difficult, in that the agent learns to acquire knowledge and establish mappings all by itself, from exemplars presented to it during training. Unlike supervised learning, unsupervised learning agents need a larger train exemplar set to adequately learn knowledge and generalize effectively on the dataset. Due to the cost of creating datasets and the quest to design learning agents that operate on less number of train patterns to yield optimal estimates, unsupervised learning becomes challenged to some extent in their implementation. Although supervised learning has advantages in terms of prior information, the unsupervised learning agent’s ability to engage in explorations so as to unearth interestingness in the dataset, is the key to machine intelligence research. For this reason, semi-supervised learning agent development approach is adopted in this book chapter and discussed thoroughly within the context of reinforcement learning.