From Trial to Policy

My name is Grigorios Agathangelidis, and my professional background is in Electrical Engineering and Software Engineering. I am not a machine-learning researcher, artificial-intelligence scientist, control-theory academic, or specialist in reinforcement learning. The ideas presented in From Trial to Policy are not the result of formal AI research programs or industrial machine-learning deployments. Rather, they emerged from independent study, practical software-engineering experience, and a desire to understand how learning systems make decisions.

This book grew naturally from the NEXUS-1 project. While From Grid to Core explored the architecture of a digital-twin platform and From Flood to Cause examined deterministic reasoning and root-cause analysis, this volume turns to a different question: how can a system learn from experience while remaining understandable to the engineers responsible for it?

My interest in reinforcement learning began not with artificial intelligence itself, but with a broader engineering concern. Modern systems increasingly rely on algorithms that adapt, optimize, and recommend actions. Yet many of these systems become more difficult to understand as they become more capable. Engineers are often asked to trust outputs they cannot fully inspect. I wanted to explore whether learning and transparency could coexist, and whether a useful learning system could remain simple enough to be understood rather than merely trusted.

Although the examples in this book are drawn from a nuclear-inspired digital twin, the underlying ideas extend far beyond that domain. Any system that must make decisions under uncertainty faces similar challenges: balancing competing objectives, adapting to changing conditions, learning from feedback, and explaining why one action was preferred over another. These challenges appear in industrial automation, robotics, logistics, finance, energy systems, transportation, and modern software platforms.

Throughout this work I have tried to connect disciplines that are often presented separately: reinforcement learning, control systems, simulation, digital twins, software architecture, observability, explainability, and operational decision support. My goal has not been to present reinforcement learning as a mysterious branch of artificial intelligence, but rather as a practical engineering technique that can be understood through familiar concepts such as feedback loops, lookup tables, state machines, and optimization.

A central theme of this book is that interpretability is not a luxury; it is an engineering requirement. The reinforcement-learning agent presented here does not produce decisions through millions of hidden parameters. Instead, it learns a finite policy represented by a readable Q-table that can be inspected, challenged, audited, and discussed. Whether this specific implementation is ultimately useful is less important than the principle behind it: that learning systems should remain understandable to the people who build, operate, and maintain them.

The algorithms, architectures, and examples presented throughout this book are educational and illustrative in nature. They are intended to demonstrate engineering concepts and implementation approaches rather than provide certified control solutions. The digital twin used for training is a simplified surrogate model, and the learned policy operates in an advisory context. Nothing in these pages should be interpreted as a validated control strategy for any real industrial facility.

What interests me most is the relationship between learning and understanding. How does experience become knowledge? How does a collection of observations become a policy? How can a decision made by a machine remain visible, explainable, and accountable to a human operator? These questions sit at the intersection of software engineering, systems thinking, simulation, and artificial intelligence, and they continue to shape my work.

From Trial to Policy is therefore not only a book about reinforcement learning. It is also an exploration of a broader engineering philosophy: that automation should remain understandable, that optimization should remain auditable, and that human judgment should remain central even when machines assist in the decision-making process.

The ideas presented here represent a snapshot of an ongoing learning journey. Reinforcement learning is a rapidly evolving field, and many approaches more sophisticated than the one described in this book exist today. Some assumptions will evolve, some implementations will improve, and some conclusions may change as understanding grows. That is not a weakness of engineering; it is one of its strengths. Progress comes from continual refinement, experimentation, and the willingness to question our own models.

If there is a single message I hope readers take away from this work, it is that learning systems do not have to be opaque to be useful. In many situations, a simple and understandable policy can be more valuable than a more powerful system whose reasoning cannot be inspected. Understanding remains a feature, not a limitation.

Above all, this book reflects a belief that curiosity, disciplined experimentation, and intellectual humility remain among the most important tools available to engineers. Technologies change, algorithms evolve, and new forms of automation emerge, but the principles of careful observation, transparent reasoning, and honest acknowledgment of uncertainty endure.

Author: Grigorios Kyriakos Agathangelidis
Greek name: Γρηγόριος Κυριάκος Αγαθαγγελίδης
Also searchable as: Αγαθαγγελίδης Γρηγόριος, Αγαθαγγελιδης Γρηγοριος, Grigorios Agathangelidis.