Reinforcement Learning (RL) is a paradigm of machine learning that primarily focuses on how to learn optimal behaviors or strategies through the interaction between an agent and its environment in a specific setting, in order to maximize some form of cumulative reward.

Unlike supervised and unsupervised learning, reinforcement learning does not learn from a labeled dataset but learns through the agent's actions in the environment and based on the outcomes of those actions (rewards or penalties).

Reinforcement learning has been widely applied in various fields, including games (such as AlphaGo), autonomous driving vehicles, robot control, recommendation systems, and more.

Through reinforcement learning, machines can autonomously learn how to make decisions in complex environments to achieve specific goals.

However, one of the current pain points in reinforcement learning research is that, to ensure that the learning rate can be appropriately set, existing algorithms require the scale of rewards or penalties to be limited.For example, for the vast majority of existing reinforcement learning problems, a default assumption is that the value corresponding to rewards or punishments lies within the range of [-1, 1]. In this case, if the algorithm receives rewards or punishments that exceed this range, it cannot function properly.

Advertisement

Inspired by scale-free online learning, Mingyu Chen, a Ph.D. student at Boston University in the United States, and his team proposed a set of scale-free reinforcement learning algorithms that ingeniously solve the aforementioned problem.

In detail, their algorithm achieves performance comparable to existing algorithms without the need to assume the magnitude of rewards or punishments.

This enhances the universality and adaptability of reinforcement learning algorithms, allowing the same algorithm framework to be effectively applied to problems of different scales and complexities, thereby expanding the scope and efficiency of reinforcement learning in practical applications.

In practical applications, scale-free reinforcement learning can be used for dynamically adjusting recommendation systems. It can process user behavior data in real-time and continuously update the recommended content to adapt to changes in user interests.Additionally, it can also be used in real-time trading analysis systems for financial markets, which can handle high-speed and highly volatile market data, and adjust trading strategies in real time.

A more important potential application prospect lies in the field of robotics and autonomous driving technology. In these application scenarios, given the criticality of safety, it is essential to ensure that robots or autonomous vehicles avoid certain specific behaviors.

This goal can be achieved through reinforcement learning, which trains by imposing penalties on undesirable behaviors.

However, if there are restrictions on the magnitude of penalties, the algorithm may take a longer time to completely eliminate such undesirable behaviors.

In contrast, this algorithm can significantly speed up the process, thereby effectively shortening the training cycle and reducing related costs.As previously mentioned, the inspiration for this project comes from scale-free online learning. Therefore, they aim to explore whether the achievements of online learning can be extended to the field of reinforcement learning.

This idea intuitively appears to be quite challenging: online learning allows people to obtain comprehensive information about rewards or punishments after interacting with the environment and taking actions, including information not directly caused by human actions.

In the context of reinforcement learning, the information people receive is usually limited to rewards or punishments directly related to the actions taken.

Furthermore, given that this project faces a scale-free problem, the fluctuations in rewards or punishments may be very large.

This requires them to obtain information about rewards or punishments in a timely manner in order to effectively adjust the algorithm parameters. In this context, their demand for information far exceeds the demand in the scaled situation.Like many research endeavors, their starting point was one of the most fundamental models in reinforcement learning: the Multi-armed Bandit problem. For this particular unscaled Multi-armed Bandit problem, there have been some research findings.

However, when they delved into the papers of these previous studies, they found that the design philosophy of existing algorithms was roughly the same as the strategies for online learning.

This similarity led to their application scope being limited to the Multi-armed Bandit problem, making it difficult to extend to a broader reinforcement learning field.

In light of this, they decided not to rely on the existing research framework, but to start from scratch and try to develop a completely new unscaled reinforcement learning algorithm.

The first progress of this research stems from an intuitive inspiration: although reinforcement learning cannot fully obtain information about rewards or punishments, it can, through some clever designs, ensure that every part of the information about rewards or punishments has a certain chance of being observed.Taking the multi-armed bandit as an example, for any strategy generated by the algorithm, a bias term can be introduced to ensure that each arm has a certain probability of being pulled. Under this circumstance, the information corresponding to each arm can be obtained with a certain probability.

Guided by this line of thought, the team developed two new algorithms for multi-armed bandits.

These two algorithms significantly optimized existing results: they not only proposed the first minimax optimal algorithm for the unscaled multi-armed bandit problem but also developed the first algorithm capable of reducing regret with high probability.

After successfully developing algorithms for multi-armed bandits, the research group shifted their focus to the general reinforcement learning problem.

The challenge they faced at this time was: in the scenario of reinforcement learning, there may not be a strategy that can ensure every bit of information about rewards or punishments is observed.For instance, in a certain reinforcement learning problem, if there exists an unreachable state, then the information related to that state cannot be obtained.

Faced with this issue, they tried various methods, but none of them led to a satisfactory conclusion.

By chance, the research team noticed a point that is often overlooked: for reinforcement learning problems, the importance of the reward or penalty information corresponding to a certain state is actually related to the reachability of that state.

In short, if for all possible strategies, a certain state is always unreachable, then they do not need to pay attention to its reward or penalty information, because this information will not help in optimizing the strategy.

Inspired by this insight, their problem was transformed into finding a strategy that can maximize the reachability of states.If such a strategy can be found, then the algorithms previously designed for the multi-armed bandit problem can be extended to apply, thereby completing the design of the scale-free reinforcement learning algorithm.

The last challenge of this project is: how to find a strategy that can maximize the reachability of states.

At this time, a paper titled "Settling the Sample Complexity of Online Reinforcement Learning" provided them with a key insight.

The study introduced a new reward-free reinforcement learning algorithm. Thus, this algorithm became a decisive supplement to this work.

It allowed them to find a tool that can be regarded as a black box, helping the research team to find a strategy that can explore every state. By combining with the previous progress, they ultimately successfully completed this research.Recently, the relevant paper was published on arXiv[1] under the title "Scale-free Adversarial Reinforcement Learning," with Chen Mingyu and Xuezhou Zhang as co-authors.

Chen Mingyu added: "I have been thinking about a very simple but often overlooked topic: how to truly automate reinforcement learning algorithms? How to make it so that humans (Ph.D. students?) no longer need to constantly manually adjust the parameters of the algorithms?"

For him, this project is more like a beginning: his short-term plan is to extend the conclusions of this paper to more general scenarios, such as Linear Reinforcement Learning (Linear RL) and Representation Reinforcement Learning (representation RL).

The long-term plan is to design reinforcement learning algorithms that require no assumptions. He firmly believes that this work can enhance the universality and flexibility of reinforcement learning algorithms and improve the scope and effectiveness of reinforcement learning in real-world applications.