Is reinforcement learning overhyped?

Translator | Li Rui

Reviewer | Sun Shujuan

OkayImagine that you are preparing to go with your friendsPlay chess, but he is not a human, but a person who does notunderstandthe rules of the game Computer program. Butthis application understands that it is committed to achieving one goal, which is to get in the game Win. ​

Because the computer program does not know the rules, the moves to start playing chess are random. Some of these tricks make absolutely no sense, and it would be easy for you to win. Suppose here that you enjoy playing chess with this friend so much that you are addicted to the game. ​

Butthe computer program will eventuallywinbecause it will gradually learn to defeat Your methods and tricksnumber. Although this hypothetical scenario may seem far-fetched, it should give you a general idea of ​​how reinforcement learning (a field of machine learning) works. Basic understanding. ​

How intelligent is reinforcement learning? ​

Human intelligence encompasses many characteristics, including acquisition Knowledge, desire to expand intellectual abilities and intuitive thinking. Human intelligence was thrown into doubt when chess champion Garry Kasparov lost to an IBM computer called Deep Blue. In addition to capturing the public's attention, apocalyptic scenarios depicting a future in which robots rule humanity have also dominated mainstream consciousness.

​

However, "Deep Blue"

is not an ordinary opponent. Playing chess with this calculationprogram is like playing chess with a thousand-year-old old Playing games, and hehas been playing chess all his life. But "Deep Blue" is good at playing a specific game, rather than other intellectual activities, such as playing a musical instrument, writing a book, conducting Science experiments, raising children, or repairing cars. ​This is definitely not intended to belittle what "Deep Blue"

has achieved

achievements. Contrary to , the idea that computers can surpass humans in intellectual abilities requires careful examination, starting with an analysis of how reinforcement learning works. How reinforcement learning worksAs mentioned in

, reinforcement Learning is a subset of

Machine

learning, which involves the concept of how an intelligent agent acts in an environment to maximize cumulative rewards. ​

Simply put, Reinforcement learningRobots are trained by a reward and punishment mechanism, they will make correct actions Get rewards, makewrong actions will be punished. Reinforcement LearningRobots don’t"think"how to take better actions,They simply make all actions possible to maximize the chance of success.

Disadvantages of reinforcement learning

The main disadvantage of reinforcement learning is that it requires a large number of resources to achieve its goals. The success of reinforcement learning in the game of Go illustrates this point. This is a popular two-player game where the goal is to occupy the maximum area on the board using your pieces while avoiding losing pieces. ​

AlphaGo Master is a computer program that beats human players in the Go game. It consumes a lot of money and manpower, including many engineers, very Rich gaming experience with 256 GPUs and 128,000 CPUs. ​

The process of learning how to win in competition requires investing a lot of resources and energy. This begs the question: Is it reasonable to design artificial intelligence that cannot think intuitively? Shouldn’t artificial intelligence research try to imitate human intelligence? ​

Support reinforcement One argument for learning is that one should not expect AI systems to act like humans, and that its use in solving complex problems requires further development. On the other hand, the argument against reinforcement learning is that AI research should focus on getting machines to do things that only humans and animals are currently capable of doing. From this perspective, the comparison between artificial intelligence and human intelligence is appropriate.

Quantum reinforcement learning

Reinforcement learning is an emerging field that is said to be able to solve some of the above problems. Quantum reinforcement learning (QRL) is a method of accelerating computing. ​

First, Quantum Reinforcement Learning (QRL) should accelerate learning by optimizing the exploration (discovering the strategy) and exploitation (selecting the best strategy) phases. Some current applications and proposed quantum computing improve database searches, factorize large numbers into prime numbers, and more. ​

Although quantum reinforcement learning (QRL) has not emerged in a groundbreaking way, it holds the promise of solving some of the major challenges of conventional reinforcement learning. ​

Business Case for Reinforcement Learning

As mentioned above, reinforcement learning research and Development is crucial. Here are some practical examples of reinforcement learning from a McKinsey & Company survey that can:​

• Optimize semiconductor and chip design, optimize manufacturing processes, and increase output in the semiconductor industry.
• #Increase factory output, optimize logistics to reduce waste and costs, and increase agricultural profits. ​
• # Reduce time to market for new systems in the aerospace and defense industry. ​
• #Optimize the design process and improve the production efficiency of the automotive industry. ​
• Grow revenue, improve customer experience through real-time transactions and pricing strategies, and provide customers with advanced personalization in financial services. ​
• Optimize mine design, manage power generation, apply overall logistics scheduling, optimize operations, reduce costs, and increase output. ​
• Increase production through real-time monitoring and precise drilling, optimize tanker routes, achieve predictive maintenance, and prevent equipment failures in the oil and gas industry. ​
• Facilitate drug discovery, optimize research processes, automate production and optimize biomethods for the pharmaceutical industry. ​
• Optimize the supply chain, implement advanced inventory modeling, and provide advanced personalized services to customers in the retail sector. ​
• #Optimize and manage networks, applying customer personalization in the telecommunications industry.
• #Optimize transportation logistics routes, network planning, and warehouse operations. ​
• Use next-generation proxies to extract data from websites.

Reflections on Reinforcement Learning

The power of reinforcement learning may be limited, but it cannot be overestimated. Additionally, as reinforcement learning research and development projects increase, so do the potential use cases in nearly every sector of the economy. ​

Large-scale adoption of reinforcement learning relies on several factors, including optimal algorithm design, configured learning environments, and the availability of computing power. ​

Original title: ​​Is reinforcement learning overhyped? ​​, Author: Aleksandras Šulženko​

The above is the detailed content of Is reinforcement learning overhyped?. For more information, please follow other related articles on the PHP Chinese website!