Introduction

Author’s note: In the context of this article, the terms “data-driven algorithms,” “deep learning algorithms,” and “AI models” express the same concept.

Before we delve into the topic of this article, we need to briefly clarify what “regulation and control” means in this context.

I believe that anyone in the autonomous driving industry is familiar with the five major modules of perception, localization, decision-making, planning, and control – although we distinguish between them in daily communication, it can be problematic if we still classify them in such a way during actual work.

The problem here lies in decision-making. Broadly speaking, decision-making includes prediction, and the division of prediction is very awkward. Prediction itself is highly related to perception, but prediction is a prerequisite for decision-making and can only be carried out after prediction has been made.

We can refer to all established facts that have already occurred as T0+. Perception is an established fact that has already occurred. For example, how the vehicle should move in the future and how others may wait are things that will happen in the future, which we can refer to as T0-.

Dividing them based on T0 has a benefit: perception only needs to be concerned with established facts that have already occurred on the road, and any inference based on these established facts can be directly placed in the decision-making part.

This classification can more closely integrate the results of prediction with the decision-making part, “weakening” the relationship between prediction and perception, and “strengthening” the relationship between prediction and decision-making. In practical work, such classification can better determine what should be predicted and how the results of the prediction can better serve decision-making. (The above content is summarized by the author after communication with Dr. Yang Wenli, CEO of Luminar Technologies.)

Based on the above logic, autonomous driving companies typically have an internal PNC team, whose main task is to work on regulation and control. If you have been working in the PNC team for a long time, you will unify prediction, decision-making, planning, and control into “regulation and control.”

Therefore, the “regulation and control” mentioned in this article refers to the same “regulation and control” in PNC teams within autonomous driving companies.## Translation in English Markdown

Next, we will combine “data-driven” with “regulation and control”.

Although “data-driven” has long been the basic consensus of the autonomous driving industry, in fact, deep learning algorithms driven by data are still mainly used in perception modules and rarely used in regulation and control. Currently, the regulation and control-related algorithms in the industry are mainly based on rule-driven methods.

Although most of the currently available public autonomous driving regulation and control algorithm-related materials are aimed at rule-based algorithms, I often see automatic driving companies require candidates who understand deep learning algorithms when recruiting regulation and control algorithm engineers. This also indicates that many companies are also “marching” towards the hybrid decision-making algorithm.

Currently, there are few articles that fully and meticulously explain the application of deep learning algorithms in autonomous driving regulation and control. Therefore, this article will elaborate on the limitations of rules and the potential advantages of deep learning algorithms, why deep learning algorithms cannot be widely used in regulation and control algorithms at the current stage, and whether future regulation and control algorithms will be based on rules or based on deep learning.

Limitations of Rules and the Potential Advantages of Deep Learning Algorithms

Rule-based algorithms have limitations

Currently, regulation and control-related algorithms are mainly based on rules, but as the level of autonomous driving increases and application scenarios continue to expand, rule-based regulation and control algorithms will expose their limitations in actual testing.

The three large corner cases that reveal the shortcomings of rule-based regulation and control algorithms are: non-deterministic scenarios, strong interactive scenarios, and strong subjective scenarios:

(1) Non-deterministic scenarios

Non-deterministic scenarios are relative to deterministic scenarios, with the simplest deterministic scenario being “red light stop, green light go”, being able to change lanes on white dotted lines, and not being able to change lanes on white solid lines. Any scenarios other than these are considered non-deterministic.

For example, in most cases, cars drive on roads with lane markings, but there is a scenario where the current road consists of three primary lanes and one auxiliary lane, with only lane markings and no greenery between the auxiliary lane and primary lanes. This means that vehicles on the auxiliary lane may merge into the primary lanes at any time. This is a typical non-deterministic scenario because most drivers do not pay attention to whether the line is solid or dotted, and you do not know when a car may suddenly appear in the vicinity, and what probability it has to change lanes.

Because the trajectories of other vehicles on the road are probabilistic, it is difficult for rule-based algorithms to handle.

(2) Strong Interactive Scenarios

The most typical example would be a crossroad where multiple objects’ routes in a scenario will interact and affect each other, so there is a certain amount of game theory involved.

In a very complex road scenario, constantly predicting the surrounding cars, and then adjusting the planning based on the results of interactions, cannot be clearly described using human rules.

(3) Strong Subjective Scenarios

Simply put, under different road conditions, people will unconsciously adjust the distance between their car and the car in front based on the environment. Because there may be a set of rules in the human brain that drives this behavior, but these rules are abstract and cannot be clearly defined, only understood implicitly. Examples include the timing of changing lanes, and reactions to other vehicles at intersections.

In some scenarios, although rule-based algorithms can imitate some of the driver’s actions and driving styles, it is difficult to solve human subjective habits with rules, and even if it is solvable, a lot of manpower is needed.Using rule-based algorithm to handle corner cases, the most challenging part is prediction. Because, in addition to predicting the trajectory of its own vehicle, autonomous driving vehicles also need to predict the trajectory of “other road users”, and there are often more than one “other road users” on the road, they will interact with each other, making it difficult to predict their trajectory.

When human drivers make predictions, they rely on “experience” and “intuition” by observing the vehicle’s historical trajectory and driver’s driving style, and can choose their own strategy based on whether other vehicles are conservative or not. However, rule-based algorithms are difficult to achieve similar predictive results as humans.

In addition, when the vehicle encounters a corner case that requires braking, the rule-based control system is difficult to ensure smooth trajectory at all times.

For example, if a vehicle ahead is traveling at a very low speed, and the autonomous driving system issues an “overtaking to the right” command, but halfway through the lane change, another vehicle suddenly overtakes at high speed in the right lane, the decision-making system will cancel the lane change decision and generate a smooth curve to return the vehicle to the original lane.

The challenge here is how to balance elements such as comfort, safety, efficiency, and feasibility of power model, and ensure that the trajectory planning path generated at this time is smooth?

Continuing with the above example, if the trajectory generated at this time cannot be guaranteed to be smooth, the vehicle will be driven in a “rough” manner, which will be uncomfortable during the journey. If the driver/passenger can see emergencies around them at this time, and the system helps them avoid accidents in time, they will think the system is reliable, even if the passenger experiences an emergency braking, they would say that the autonomous driving system “helped me avoid an accident”.However, in reality, in most cases, the system has already detected the problem and made decisions to prevent it in advance. The issue is that most drivers/passengers do not realize that a seemingly normal lane change could have caused an accident or may not even be aware of oncoming vehicles. Therefore, if the system cannot generate a smooth curve at this point, drivers/passengers will not think that the system “saved them from an accident”, but instead complain that the system is “too rigid” and the lane change feels “jerky”.

Although such problems can also be solved by rule-based algorithms, the process is very complicated. Moreover, although theoretically speaking, some niche scenarios can also be solved by rule-based algorithms, engineers find it difficult to ensure that these complex codes can solve the problem without using too many computing resources in actual work. While it may be easy to run these codes on offline systems, can the on-board chip of a car support such a high level of computing power? Solving these problems highly challenges the coding level and engineering ability of the engineers.

In addition, some engineers may have written thousands of lines of code and then forget what they have written. The more rules there are, the more confusing the logic becomes. When a new demand is added in the “future,” it will become difficult to maintain the system.

2. Deep learning algorithms can improve the human-likeness of the system in suitable scenarios.

The limitations of rule-based algorithms are not limited to the ones mentioned above. If we look at all problems comprehensively, we can generalize the limitation of rule-based algorithms as “low human-likeness” of the system.

Human-likeness refers to whether the system acts like an “experienced driver” during the driving process. One rule-based algorithm expert of Momenta says:

“The core ability of human-like experienced drivers is the ability to adjust to changing scenarios. In contrast, rule-based systems require a lot of rules to differentiate scenarios, set up a large amount of environmental parameters and cost functions, and constantly seek optimal solutions, which makes real-time operation mathematically difficult.”Currently, autonomous vehicles that are able to be tested on the road resemble more of a “novice driver” in most situations, while the biggest difference between a “novice driver” and an “experienced driver” is that the latter has stronger predictive ability and is better at adapting to random scenarios in long-tail situations.

In comparison, although most “novice drivers” are also safe, they often drive very “naively” on the road, while “experienced drivers” can perform various operations more smoothly and efficiently while ensuring safety. This difference is called “parole experience.”

Essentially, “novice drivers” rely more on “driving school-based teaching materials” on the road, and their decision-making and planning methods are more like if else (rule algorithms), while “experienced drivers” rely on experience and subjective judgment when making decisions, and their decision-making and planning methods are relatively abstract, more like “deep learning algorithms”.

Therefore, by introducing deep learning algorithms into the regulatory control process, learning as much as possible from human driver experience, automatically extracting features and finding rules, it is possible to improve the “human likeness” of the regulatory control algorithm.

The data engine mentioned by Tesla at this year’s AI Day is to improve certainty by inputting data into neural networks to better solve real-world prediction problems.

The head of regulation and control algorithm at Light Sail Intelligent Navigation believes that predicting (judging the future trajectory of other vehicles) is the most suitable and the first to be widely used for deep learning algorithms. He said: “The prediction algorithm of Light Sail Intelligent Navigation uses deep learning.”

Regarding specific scenarios, he believes that scenarios that rely heavily on subjective judgment of humans require more guidance from deep learning algorithms.

The regulation and control algorithm expert of Pony.ai stated that Pony.ai has applied AI models at multiple levels.>”For example, traditional AI models (either deep learning or traditional machine learning classification models) can be used for decision making, deep learning can be used for trajectory generation to guide path planning, and end-to-end reinforcement learning can be used to provide instructions for motion planning (the final trajectory planning).”
>

This expert also believes that structured high-speed scenarios may be more easily solved by traditional algorithms, while non-structured urban scenarios may offer greater opportunities for AI models to excel.

ZhiTu uses reinforcement learning algorithms for the decision-making stage in highway ramp scenarios.

Dr. Yang Wenli, CEO of Momenta, believes that the importance and applicability of deep learning algorithms for different modules of autonomous driving are not exactly the same, from perception – fusion – prediction – decision – planning – control. “The earlier the stage, the better the effect of deep learning algorithms.”

Dr. Yang Wenli said:

“The perception system has a large amount of data processing, but allows for errors and has some tolerance for accuracy and recall (because fusion and decision-making can correct perception errors), so deep learning methods are more appropriate.

“For this decision-making stage, rules and learning each have their own strengths, so we use a hybrid architecture, using deep learning methods to improve performance and rule methods to ensure safety.

“The control system has a small amount of data, with strong reliability requirements, absolutely no room for error, and the use of deep learning algorithms cannot only improve performance, but can also introduce uncertainty and unreliable factors, so it is not suitable to use deep learning algorithms. In addition, control, as a fundamental technology, has a history of 200 years, and its mathematical proof is already perfect, so there is no need for deep learning.”Moreover, in case of errors in the planning algorithm, the control algorithm also needs to act as a safety net. In classical PID control algorithm, the control layer can reject the “erroneous command” issued by the planning layer to ensure safety. For instance, if the planning layer issues a command for the vehicle to drive at a speed of 120 km/h but its hardware cannot support such a speed, the control algorithm can perform a check and reject the command directly to ensure the normal operation of the vehicle without overturning.

In order to ensure reliability in case of all possible errors, the control algorithm must use rule-based algorithms.

Therefore, when people mention “the application of deep learning algorithms to guidance and control”, the term “guidance and control” mainly refers to prediction, decision-making, and planning, but not control.

In addition to solving practical technical problems, there is another significant advantage of using deep learning algorithms, which is to take full advantage of GPU or neural network specific processors to reduce the CPU computational power consumption.

To achieve certain functions, in the early stages of autonomous driving development, complex rule-based code was used, which consumed a lot of CPU computational power. However, the design of on-board chips tends to enhance deep learning computational power, and the CPU part computational power is limited, so engineers must find ways to save it. By replacing a large number of complex rule-based algorithms with deep learning algorithms, a considerable amount of CPU computational power can be saved.

According to the control algorithm experts of Landing.AI, the NOA production cars in urban areas generally use embedded chips, and the CPU is based on ARM architecture. The CPU computational resources cannot be compared with that of the L4 Demo car based on X86 architecture. Therefore, if the suppliers want to achieve a driving and riding experience close to L4, it is necessary to try to introduce deep learning algorithms to the guidance and control process and put the load on the deep learning-specific processor.

At present, deep learning algorithms cannot be applied to guidance and control algorithms on a large scale.Since deep learning algorithms sound so “amazing,” why do rule-based algorithms still dominate in practical applications rather than deep learning algorithms?

This is because deep learning algorithms also have many shortcomings. It is precisely these shortcomings that lead to the fact that in the short term, the effect of deep learning algorithms is not as good as rules.

1. It’s hard to determine whether human drivers’ behavior is “worth learning.”

In an ideal situation, the more data used for training, the better the effect of deep learning algorithms. However, in some cases, as the amount of data increases, the effect of the algorithm may decrease. Why is this?

Deep learning algorithms need to first learn a large amount of human driver behavior before they can be applied to rule-based algorithms. However, the problem is that the driving styles of different drivers are different, and the same driver’s driving style is not completely consistent under different circumstances. So, how can we ensure that human driver behavior is good data worth learning? For example, for some lane changing behaviors at intersections, we cannot cross the solid line to change lanes during normal driving. However, many drivers often like to cross it.

The regulation algorithm expert at Pony.ai said:

“Human drivers also have many bad habits. If we blindly pursue anthropomorphism and ignore safety, comfort, and efficiency, which are the main optimization goals of autonomous driving systems, we will reverse the order of importance. In fact, even if AI models are used to make up for the shortcomings of traditional algorithms, they should also be improved around these three dimensions.”

Therefore, in many cases, only by pre-cleaning data, can the effect of deep learning algorithms be good. However, how to ensure the efficiency and quality of large-scale real-time data cleaning is still a problem.## Translated English Markdown Text

Regarding this issue, some companies adopt the practice of setting rules to evaluate drivers’ performance and filtering out those who appear to perform poorly. However, using rules to “constrain” deep learning is overly simplistic and may result in the deletion of useful driving behavior data, much like “throwing out the baby with the bathwater.”

2. Deep learning algorithms may struggle to adapt to changing demands compared to rules-based algorithms

An autonomous driving company has a deep learning algorithm that specifies a minimum safe distance of five meters from the vehicle in front. One day, the product manager proposes a new requirement to increase the minimum safe distance to ten meters.

This situation highlights a problem with deep learning algorithms: engineers would need to retrain the model at significant cost to make changes according to new requirements. However, using rules-based algorithms would be much simpler – a few lines of code could solve the problem.

Furthermore, if an accident occurs, it is difficult for deep learning algorithms to quickly identify the problem like rules-based algorithms can.

3. Deep learning algorithms may not be as effective as rules-based algorithms

According to experts in rule-based algorithms at Qingzhou Intelligent Navigation:

“If AI models are introduced without a well-established data infrastructure, it is difficult to achieve the expected result.”

For example, it may be challenging for an AI model to achieve the same results as simple testing and calibration of an adaptive cruise control system. The main reason being the lack of diversity of data under different follow-up scenarios, which makes it difficult to train a model that performs at or above the level of rules-based algorithms.

Yimuyun Intelligent Driving’s experts in rule-based algorithms also agree:> “Currently, AI models have not completely outperformed the rule-based models. Although attempts have been made to simulate human cognition with AI models, the results are not satisfactory. This is due to the fact that the model’s architecture, input/output definition, and data preparation have not yet reached the level of perceptual models. Everyone is still exploring, but I believe there will be significant breakthroughs within a year.

4. Deep learning algorithms still need rules as a backup

In the process of driving on the road, it is impossible to completely avoid errors in decision-making algorithms, even for experienced drivers. The same applies to decision-making algorithms based on deep learning models.

Therefore, when designing systems, developers should consider the possibility of “decision errors” and be mentally prepared to “accept” this potential fact.

In response to this, Pony.ai has adopted a hybrid decision-making architecture. The CEO of the company, Dr. James Peng, said, “Deep learning algorithms are good at learning human driving experience, and are mainly used to improve performance. Rules designed by programmers have been fully tested and demonstrated, and have higher priorities to ensure safety.”

Going deeper, the planning layer can also reject “incorrect commands” from the decision-making layer. For example, the speed and size of obstacles should be within a certain range (such as speeding cars that “take off” at 100m/s are not allowed); or, the decision-making layer may give the command “change lanes to the left”, but there are only cars on the left, and changing lanes at this time would result in no perfect space planning solution. In this case, planning can directly “reject the command”.

The planning algorithm that rejects “incorrect commands” from the previous layer can only be rule-based.

Based on the above, we can see that in scenarios where deep learning algorithms can be used, rule-based algorithms still play an essential role as a backup.

5. Introducing deep learning algorithms means a large amount of code needs to be replaced.There is another challenging issue for autonomous driving companies: if mixed regulation and control algorithm is used, a significant amount of code needs to be replaced, and this inevitably requires a large number of skilled personnel.

Experts in regulation and control algorithms at QZSS say that traditional rule-based systems were not designed with interfaces related to deep learning, and the code library has been relatively solidified after years of development. The cost of rebuilding the code for introducing deep learning is very high.

“But QZSS has considered the extensive use of AI when building the decision-making and planning algorithm stack since 2019 and has proactively considered the situation after the model is introduced at various interface levels, and made targeted architectural adjustments”.

According to a senior expert who once served as the regulation and control algorithm leader at a top Robotaxi company, introducing deep learning into regulation and control would require a massive amount of code rewriting because it involves multiple systems such as online, offline, and simulation.

As a writer, I’ve heard numerous complaints from users in the internet industry about why some software developers always choose to patch rather than restructure complex systems. This example is similar to the problem faced when using mixed regulation and control algorithms to replace rule-based regulation and control algorithms – the benefit of restructuring is low.

You rewrite the code, but the improvement in the system’s operating level is limited, and you have to pay a massive cost for it. This kind of “hard work may not necessarily bring satisfaction” is a “heavy burden that cannot be borne” for any enterprise. This concern also makes autonomous driving companies have a psychological aversion to mixed regulation and control algorithms.

In fact, after communicating with regulation and control algorithm leaders/engineers of several autonomous driving companies, I found that different from the “scene” when deep learning neural networks were applied to perception in 2014-2015, everyone is generally in no hurry to apply deep learning algorithms in the regulation and control link.

It can be said that everyone is still researching and practically applying deep learning algorithms with an attitude of “optional.”The self-driving companies know that the “hybrid rule-based and deep learning algorithm” is the future, but there are currently few scenarios where this algorithm can be used, so they decided to recruit some experts to “research on their own”.

Therefore, currently, most self-driving companies only use the hybrid rule-based and deep learning algorithm to deal with a few scenarios such as ramps, intersections, etc., and for the corner cases that have already been solved by rule-based algorithms, they will leave it unchanged as long as there is no “intolerable” problem discovered.

Moreover, rule-based and control algorithm experts bluntly point out that whether to use hybrid rule-based and deep learning algorithm or not depends on demand, which is an engineering issue.

According to the above logic, currently, the attitude of self-driving companies towards hybrid rule-based and deep learning algorithms is “optional”, and the application scope of this algorithm in the future also needs to be questioned.

Rule-based or Deep Learning Algorithm-based?

We continue to explore the logic from the last section.

Many experts believe that no matter how mature the development of deep learning algorithms and hybrid frameworks is, rule-based and control algorithms will always account for the majority, with deep learning algorithms only playing a supporting role.

Their key reason is that deep learning algorithms have the characteristic of “inexplicability”. When using deep learning algorithms for prediction and decision-making, they follow probabilities instead of strict causal relationships. Therefore, it is difficult to trust them as the “primary responsible party” for rule-based and control.

As Dr. Yang Wenli, CEO of Momenta, said:

“Deep learning is a black box. The ‘scene’ that humans understand and the ‘scene’ that deep learning understands may not be the same. There was a diss example of AI before. Changing a few pixels in the image caused AI to recognize it incorrectly. If you put some black tape on traffic signs, Tesla won’t be able to recognize them. Unless it is pixel-level ‘consistency’, what people regard as the same ‘scene’ is highly likely to be ‘different scenes’ in deep learning’s perspective.”

“Deep learning algorithms are based on input-output fitting, which can obtain better human-like performance from the characteristics of input-output. However, it is difficult to understand the internal logic relationship of driving. “In order to reduce the possibility of errors in decision-making using deep learning algorithms, we have adopted a mixed decision-making architecture.”

CTO of a commercial self-driving car company and a regulation and control algorithm expert of a new force in the automobile industry hold similar views.

At a technical exchange event at the end of September, Dr. Huang Chang, CTO of Horizon, also talked about this issue, but Huang Chang believes that in the mixed regulation and control algorithm, deep learning will be the mainstream, and rules will only be “necessary assistance” – to avoid obvious errors that can be interpreted semantically.

Huang Chang explained that in his experience, for a polished and well-tuned rule-based system that has been in development for many years, introducing deep learning into only one local module in a hybrid manner would have limited or even no improvement, “because other modules are still designed based on rules. In this large framework, the cost of adapting other modules to this new module is high. The validation cycle is so long that you are not patient enough to complete it.”

The regulation and control algorithm responsible person of a certain Robotaxi company and the regulation and control algorithm expert of a certain Liangzhou also believe that the future regulation and control algorithm is mainly based on deep learning, and rules are auxiliary.

“In Liangzhou Zhihang, in the prediction link, deep learning algorithm related operations (including preprocessing, model inference, and post-processing) account for nearly 95% of the overall module operation time. In the decision-making and planning links, deep learning algorithm operations account for about 30%.”

“It’s not easy to quantify, but overall, in the regulation and control algorithm, deep learning algorithm should account for 50%-60%.”He believes that ultimately, 80%-90% of the operational time of regulation algorithms will be spent on deep learning algorithms. Unlike Dr. Yang Wenli’s mention of “no need to use AI in the lower-level control algorithm,” he believes that control algorithms are mainly based on rules, but deep learning models can also be used to optimize the control parameters to increase algorithm adaptability.

He pointed out:

“If there were disputes in the industry before about whether ‘AI is the main’ or ‘rules are the main’, it is understandable, but the information released by Tesla at the latest AI DAY has shown that Tesla uses a large number of deep learning regulation algorithms which perform very well, which indirectly reflects Musk’s ‘first principle’ can also apply to deep learning-based regulation algorithms. If the human brain, as a complex neural network, can learn to drive through a lot of experience, then regulation algorithms can also use deep learning to achieve better performance and results.”

There is a great controversy in deep learning algorithms, which is their “unexplainability”. Therefore, if a certain scenario changes from “unknown” to “known”, do we need to add rules to determine it in the algorithm?

To this question, a regulation algorithm expert from a Robotaxi company believes that it is “feasible” to determine the uncertainty of AI’s scenario identification through rules; while Huang Chang and Qingzhou regulation algorithm experts believe that it is “not feasible” and “not necessary”.

Huang Chang said:

“We cannot force this system to be ‘explainable’ – because the so-called ‘explainable’ means to constrain it with rules, or even completely rely on rules like expert systems. The result is that the system runs well in City A, and when it goes to City B, it still needs a large number of engineers to debug it; and it makes those complex corner cases that cannot be described by human rules ‘unsolvable’.”可见，强求系统一定要 「可解释」，这就又回到了本文第一节所提到的 「规则的局限性」 上了。

黄畅认为，在将深度学习算法引入规控时，我们不必过分纠结于系统是否 「可解释」。

「我跟你交流，我们之间，我能够充分理解你，并能推理你的整套逻辑，但对你底层的一些直觉、一些隐藏的特质，我却是不知道的（甚至你自己也未必知道），但这并不影响我们能建立起信任感、能顺畅地交流。同理，我们也不必用是否‘可解释’去限制去探索深度学习算法的应用边界。

**「我前段时间看了一篇帖子，说深度学习和经典的统计学之间的区别在于经典统计学依然尝试用规则模型的方式让整个系统变得‘可解释’，但深度学习突破了这一点，它甚至非常极端去优化最终的目标。

「人会犯错，规则系统会犯错，数据系统也会犯错，但是如果在广泛统计学意义上，当数据规模足够大的时候，深度学习算法会比规则系统强一个数量级，因此，我们可以放心地使用它。

**「在这个基础上，把环境模型这样的东西可视化，让人可以理解，再在规划的最后一个阶段引入少量的必要的规则去约束深度学习算法，这样就足够了。」

黄畅认为，在算法 2.0 时代，AI 有能力去完成自适应，因此，自动驾驶公司可能要设计一些 「在算法之上的算法」，使得它能够帮助基础的算法更好地在应用场景中去迭代、去适应。

轻舟规控算法专家认为，笔者提到的 「当一个场景从未知变成已知的时候，就用规则把应对方案给‘确定’下来」 的设想在实践中很难行得通，因为，现实中的场景往往是动态调整的 —— 起初可能是清晰的，但变着变着，就 「模糊不清」 了，因而很难用规则将其描述清楚。Regarding the “unexplainable” nature of deep learning algorithms, the response is: “Many people say that deep learning is ‘unexplainable,’ but in fact, if you carefully design your deep learning algorithm, it can achieve ‘explainability’ and even stronger than rules.”

He gave an example: when a self-driving car turns right at an intersection and needs to avoid a car that is going straight, it is difficult to use rules to exhaustively enumerate what actions the self-driving system should take under different circumstances such as the speed of the car going straight and the distance between the two cars. However, if deep learning algorithms are used, as long as the position and state of the car and the potential obstacle (the car going straight), as well as some relevant environmental and historical information are provided, the trained neural network can calculate the probability of “you give way to me” or “I give way to you,” which is the same decision-making process as that of a human driver. “Probabilities are easier to explain than rules.”

At this point, the author thought of a seemingly unrelated topic, but has a deep connection with the main theme of this article-what is human “intuition”?

Previously, many people who studied enterprise management often said that entrepreneurs rely on intuition rather than research when making important decisions. Few people can explain what this intuition is, but the author thinks that this “intuition” is actually data-driven “deep learning algorithm.”

The so-called “accurate intuition” is not a talent. Behind it is accumulated information and cognition based on this information. Similar to big data analysis, there is correlation, but no causal relationship between conclusions and information, or the causal relationship is not direct or obvious, so it is difficult for humans to explain it logically, so it is called “intuition.”

In this sense, we really don’t have to insist that deep learning algorithms must be “explainable.”

However, the experts in the rule-based algorithm control at Momenta believe that even if a large number of AI models are used for cognition, they must be scenarized and have explicit and explainable “intent and meaning”. “This is our principle.”

The expert stated:

“The strength of deep learning lies in its ability to adapt to various complex environments and quickly provide a solution. However, its black-box nature leads to incorrect solutions at times. On the other hand, rule-based algorithms are based on the knowledge of what it knows and what it doesn’t know. They may have a lower probability of providing a correct solution in unknown complex scenarios compared to deep learning methods, but the inability to provide the correct answer does not affect the evaluation of the answer after seeing it. Therefore, AI is used as the main method for problem-solving, while rules are used to verify the results. “

The regulation and control algorithm expert of Pony.ai believes that at the current stage, it is challenging to ensure the safety of deep learning algorithms. Therefore, in the foreseeable future, traditional algorithms and deep learning will be combined.

The regulation and control algorithm expert stated:

“We believe that L2 can potentially achieve the state of AI as the main method, as AI does not need to handle all corner cases – after all, drivers can provide the final safety guarantee. Tesla’s presentation at the AI Day also proves this point.

“However, it will be difficult for L4 algorithms to rely solely on AI as the primary method in the foreseeable future, mainly due to its explainability and handling of corner cases with security risks. However, with the advancement of technology, we believe that the AI portion of algorithms will increase.

“In fact, traditional regulation and control algorithms are not simply a stack of rules. Good traditional regulation and control algorithms are mostly mathematical models abstracted from the induction and summary of scenarios – modeling planning problems with different mathematical tools and solving them through optimization. They have the ability to address the essence of the problem and fundamentally ensure safety within the theoretical scope.”> “Therefore, the introduction of deep learning does not involve rewriting a large amount of code. Instead, it is more about how to combine traditional algorithms with deep learning to ensure the safety of driving using traditional algorithms. It can be said that traditional rule-based algorithms and AI models are not a simple substitute relationship, and algorithm development is also an iterative process. There is no need to spend a lot of cost to replace rule-based code with AI models immediately. It is a gradual process.”

Meanwhile, an architect from a major automaker suggested running rule-based algorithms in the main system and running deep learning algorithms in shadow mode. When the training effect of the deep learning algorithm exceeds that of the rule-based algorithm, then the deep learning algorithm can be deployed to the main system.

Attached. Two cases of deep learning algorithms applied in autonomous driving control

At the end of July, Juzhen Intelligence Driving held a technology salon themed “Data-driven Rule-Based Algorithms” in Suzhou High-speed Railway New City. At this salon, experts from Qingzhou Intelligence and IMa shared some simple examples of how deep learning algorithms can be applied in autonomous driving control. With the consent of the experts concerned, we share these two cases as follows:

1. IMa: Reduce fuel consumption through deep learning algorithms

The application of data-driven decision-making by IMa has also evolved from being primarily based on rules to a mixture of rules and data-driven decision-making. Cui Dixiao, the chief scientist of IMa Technology, mentioned that they believe the timing of lane-changing is very suitable for using deep learning algorithms to assist in decision-making.

Because trunk logistics have high requirements for timeliness and fuel consumption, inappropriate lane-changing timing will lead to a decrease in the percentage of using the automatic lane-changing function by drivers. On the other hand, if the rule-based lane-changing algorithm is too conservative and only triggers the “lane-changing” command when the speed of the front vehicle is too low, it will cause more subsequent braking, resulting in a loss of time efficiency and fuel consumption. In order to reduce fuel consumption, the autonomous driving system collects data to observe how drivers make decisions and then researches more human-like lane-changing decisions.

This solution has gone through two stages.# Stage 1: Changing from driver-initiated to system-suggested lane change decision.

During this stage, engineers continuously study why drivers make certain lane change decisions at certain points and develop rules for lane change suggestions. After the system suggests a lane change, if the driver allows it, the system considers the decision appropriate. If the driver disagrees, they can choose not to follow the suggestion.

Stage 2: Improving fuel efficiency through human-like decision-making using deep learning algorithms.

According to predictions and control experts at QZ Intelligent, the current control algorithm has left ample interfaces for deep learning, and they are actively developing deep learning control algorithms. The ultimate goal of QZ Intelligent’s overall control algorithm framework is to achieve AI First by having the model output most behaviors and then using rules to complement it (rules mainly deal with safety-related issues).

QZ Intelligent believes that control capabilities will be an important manifestation of NOA function differentiation in cities. Currently, QZ Intelligent has begun to try to use AI to handle as many aspects of control as possible in the city NOA plan, enabling the system to make decisions that are “human-like” in most situations and “beyond human” in extreme situations.

For example, QZ Intelligent already has dedicated deep learning algorithms to handle certain critical planning scenarios encountered when using city NOA, such as lane change timing selection, ramp merge timing selection, and intersection games.

(Note: HTML tags are preserved in the above English text.)To ensure the predictability and controllability of deep learning algorithms, QZ Zhihang uses validation mechanisms for input and output data to ensure that input data meets the current design requirements. Rule-based methods are used in the input layer to select scenarios suitable for deep learning algorithms, ensuring the validity of the model inputs and limiting the model to the problem scope designed. Rule-based methods are also used in the output layer to validate outputs and ensure that they achieve the desired effect. If the model encounters scenarios that it cannot process well, it can automatically record the time point and scenario-related information, and update the model through data loops, making the model more versatile and adaptable to more scenarios.

Currently, at QZ Zhihang, the technology of generating “possible future trajectories” for vehicles through deep learning (such as imitation learning or reinforcement learning) is already quite mature. The model is trained by collecting large amounts of expert-driven trajectories, and in 90% of cases, the generated trajectories do not require further modification. However, in 10% of cases, the trajectory generated by deep learning may not be effective, and may even pose a safety risk. In this case, a framework is needed to correct or modify it. QZ Zhihang uses its domestically initiated spatiotemporal joint algorithm (a numerical optimization-based algorithm) to fine-tune the trajectories generated by deep learning.

Experts in QZ Zhihang’s algorithm team say that spatiotemporal joint planning algorithms are more suitable for use with deep learning algorithms than traditional horizontal-vertical separation. This is because the trajectory generated by deep learning is actually a spatiotemporal joint trajectory, which includes position, speed, direction, and time information. Therefore, spatiotemporal joint planning algorithms can be better combined with the results produced by the model, and can seamlessly provide security protection to deep learning algorithms.

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.