Open-Loop VS Closed-Loop
Introduction
In control theory, there are two types of control systems: open-loop control systems and closed-loop control systems. To evaluate a control system, there are two types of evaluation systems: open-loop evaluation systems and closed-loop evaluation systems.
In this blog post, I would like to quickly discuss the differences between open-loop and closed-loop control systems and evaluation systems and their applications in autonomous driving.
Open-Loop VS Closed-Loop Control Systems
An open-loop control system is a type of control system in which the process output is not fed back to the input. In other words, the control action is independent of the process output.
In contrast, a closed-loop control system is a type of control system in which the process output is fed back to the input. In other words, the control action is dependent on the process output.
For example, consider adjusting the temperature of a room to certain degree using an air conditioner. In an open-loop control system, the air conditioner will have to know the some information about the room before action, such as the room temperature, room size, and the desired temperature, make some assumptions, compute the power and the time needed to turn on the heater or cooler of the air conditioner. In a closed-loop control system, the air conditioner will only have to know the desired temperature before action, and the air conditioner will adjust the power and the time needed to turn on the heater or cooler of the air conditioner based on the room temperature it monitors.
As a result, an open-loop control system can be somewhat simple to implement, but it may not be accurate and reliable if the information it collected and the assumptions it made before action are inaccurate or have changed during the process. In contrast, a closed-loop control system can be more complex to implement, but it can be more accurate and reliable because it can adjust the control action based on the process output it monitors.
An open-loop control system can be used in some simple and deterministic processes, such as turning on a light switch, but it may not be suitable for some complex and stochastic processes, such as autonomous driving.
Open-Loop VS Closed-Loop Evaluation Systems
Somewhat similar to the open-loop control system, an open-loop evaluation system is a type of evaluation in which the control system is evaluated against some completely deterministic processes that the control system has no effect on.
In contrast, a closed-loop evaluation system is a type of evaluation in which the control system is evaluated against some processes that the control system has some effect on.
For example, consider evaluating a language model which generates a sequence of tokens. In an open-loop evaluation system, given a sequence of tokens generated from human natural language as ground truth, the language model will only be evaluated to see how possible it can generate the same sequence of ground truth tokens and such evaluation criteria is usually simple. Even if the language model is autoregressive, during the autoregressive token generation, the newly generated token will not be fed back to the autoregressive language model for generating the next token. In a closed-loop evaluation system, the language model can generate completely novel sequence of tokens. Such evaluation, however, is usually more complex and requires more sophisticated evaluation criteria.
There are some more realistic non-technical examples. A student examination consisting of multiple-choice questions is an open-loop evaluation system, because there are ground truth answers for each question and the student has no effect on the evaluation process. A sports game, however, is a closed-loop evaluation system, because the players are actively interacting with the evaluation process.
It should be noted that an open-loop or closed-loop evaluation system usually does not care whether the control system being tested is open-loop or closed-loop. Both open-loop and closed-loop control systems can be evaluated under the same evaluation system.
Autonomous Driving
In autonomous driving, usually the control system is closed-loop because the driving environment is just too complex and stochastic and an open-loop control will not be accurate and reliable enough to operate a vehicle safely. Imagine an autonomous vehicle takes the passenger from location A to location B, because it is impossible to predict the traffic, the weather, the road conditions, etc., during the commute, an open-loop control system will not be able to operate the vehicle safely from location A to location B.
Evaluating an autonomous driving system, however, can be open-loop or closed-loop. In an open-loop evaluation system, the autonomous driving system can be evaluated against some completely deterministic driving scenarios that the autonomous driving system has no effect on, usually pre-recorded driving courses. For each scene, the control system can produce control signals and evaluate the control signals against the ground truth control signals. In a closed-loop evaluation system, the autonomous driving system can be evaluated against some driving scenarios that the autonomous driving system has some effect on. The closed-loop evaluation system environment can be a simulation environment where the autonomous driving system can interact with the environment such as other vehicles, pedestrians, traffic lights, etc. It can also just be a real-world driving environment where the autonomous driving system is operated on road in the real world.
Conclusions
Open-loop versus closed-loop is usually about whether there are feedback in the system process.
References
Open-Loop VS Closed-Loop