The title of this article is ripped straight from the book “Artificial Intelligence” by Michael Wooldridge which is what I am going to be discussing in the article today. This is a short 50 page or 1-2 hour read from the Ladybird Expert Book series.
The main reason I read this book is because it is on the recommended reading list for the Oxford University Department of Computer Science and it was a joy to read – especially being so short and sweet.
This book details and outlines all of the main ideas and periods within the history of Artificial Intelligence leading up to the present recounting all of the different approaches and attempts to go down the path towards a true AGI (Artificial General Intelligence).
My favourite part about this book is the manner in which each approach is presented with its each intuitive understanding and biological perspective which allows the creation of links between different disciplines within the artificial intelligence community. Whilst pondering over it yesterday I ended up attempting to draw links between the Knowledge-Based AI systems and possibilities of modern reinforcement learning systems augmenting and rectifying many problems with the prior approach – particularly the need to hard code every argument. Those systems can hardly be called intelligent, as they are simply given lists of conditions to consider but if an algorithm were able to learn these conditions (such is as done in reinforcement learning scenarios through the policy) then perhaps that Knowledge-Based system may be truly intelligent.
By understanding the past there are many connections and developments that can now be made with advances within fields of machine learning and increases in computational power and distributed computing techniques.
Prominent researchers at OpenAI believe (according to their spinning up documentation) that many new developments and advancements in fields come by combining interpretations of modern techniques and slightly dated ideology – which did not work very well when first discovered.
An example of this is TRPO or Trust Region Policy Optimisation which uses an idea proposed by Natural Policy Gradients – an idea which was found to not perform as well as suggested. However, by combining newer Policy Gradient ideas with the Natural concepts OpenAI researchers such as John Schulman have managed to develop some of the most prominent and powerful reinforcement learning algorithms present today.
The problems, their solutions, and their thinking points are mostly described in this book to give the reader a perspective and an introduction to that perspective whilst encouraging them to think more deeply on their own and hopefully delve into the area and research about it. It fosters interest within these disciplines and I know it certainly did for me.
As made evident by this book, many initial AI ideas and advances were halted and limited by the computing resources of the time as well as the incapability of machines to perform certain tasks and some of those ideas and advances may possess key knowledge and information regarding different directions and pathways to follow in artificial intelligence research. A lot of them focus on robust ideologies about the working of human brains and how we learn which is something that (with today’s deep learning architectures and computational power) can finally be explored and applied to the frontier of AI research.
Books like these from experts in their fields are truly amazing stimulants for the human brain to provide an insight into certain areas of study and stimulate the mind of those already seeking to be immersed in the topic like myself. For anyone considering or who is studying artificial intelligence or computer science in general, reading this book would be extremely beneficial to get a look into one of the most futuristic sci-fi areas of study which is growing strongly in popularity and in power to either concentrate on or learn something from. There’s always something crazy and amazing about AI.
The Neural Network 