III. RIESGOS DERIVADOS DE LOS ACTIVOS QUE RESPALDAN LA
2. ACTIVOS SUBYACENTES
Subtly different to the Human Driver models, speed control algorithms have been developed to achieve high driving performance. Several algorithms have also been produced in an attempt to create a perfect driver that delivers high quality driving performance, rather than to recreate the human driver. An application of such an algorithm would be an adaptive cruise controller for example. This is where the vehicle wishes to maintain a safe distance from the vehicle in front whilst also maintaining passenger comfort by limiting acceleration for example. These algorithms fall into similar categories as the human driver models.
Adaptive Cruise Controllers (ACC) operate as an extension of standard cruise controllers. ACCs adapt to the environment around them in order to maintain safe conditions. To do this, many ACC equipped vehicles utilise a radar or laser system to detect any vehicles around them. However, with many vehicles following each other stability issues can occur (Rajamani, 2011). It is possible for an individual vehicle to be stable, but for a string of vehicles to be unstable. This is due to the delay in response of each vehicle, and the lack of information available on what vehicles further
1.4 Literature Review: Autonomous Speed Control 19 down the chain are doing. An analysis of this phenomenon results in conditions on the properties of the ACC control system (Swaroop, 1997).
A non-cooperative car following control law was proposed by Chien and Ioannou, (1992) that was free of oscillations and the ‘slinky effect’. This used relative distance, relative velocity and relative acceleration between the host vehicle and the vehicle immediately in front. The key to eliminating oscillations in this case was the introduction of a safe following distance as a function of speed.
Sivaji and Sailaja, (2013) proposed an adaptive cruise controller that utilised a PID controller in a similar method to the benchmark Chandler model mentioned earlier (Chandler et al., 1958). However, in this model, the ‘driver’ receives inputs from the speed of the host vehicle, distance between the two vehicles and a target time headway (THW).
It is possible to customise a PID controller by introducing gain scheduling, (Shakouri et al., 2011). This allows different gains to be used at different speeds, meaning the controller can perform better over a range of speeds.
Yanakiev and Kanellakopoulos, (1996) suggest a cruise controller for use specifically in HGVs. A specialist controller is needed as an HGV is often very power restricted, and so a drive profile is determined by amount of available power, rather than just the driver’s desired speed. A linearised HGV model is used and a couple of control techniques explored: standard and adaptive PID controllers. It was found that for good control characteristics, especially when in truck platoons, aggressive control actions were needed. This was due to the much heavier mass of HGVs compared with cars. However, the heavier mass reduces the discomfort of passengers by restricting acceleration and jerk of the vehicle. A non-linear adaptive PID controller was found to have the best characteristics of those tested, but did require the most tuning to the individual vehicle, even though steps were taken to extend the operating range of the controller.
In a different approach at ACC design, the standard drawbacks of a Dynamic Programming controller (such as slow convergence) were reduced by the implementation of what was called a Supervised Adaptive Dynamic Programming (Zhao and Hu, 2011) – see Figure 1-8. Machine learning, like this, can offer the potential for excellent control in a wide range of conditions, but the large training times required can be a major disadvantage. In this example, the control was split into two controllers: an upper and a lower controller. The upper controller studied the environment and used a Reinforced Learning (RL) approach to decide a desired acceleration. The lower controller utilised a fuzzy logic control algorithm to control the pedal and brake inputs to the host vehicle. In comparisons with a standard PI controller, it was noted that the SADP controller performed well, especially in an emergency stop situation, where it outperformed the standard ACC.
20 Introduction and Literature Review
Figure 1-8 – Supervised Adaptive Dynamic Programming Control System (Zhao 2011) As well as being used to model a human driver, Model Predictive Control has also been used in speed control algorithms for ACC’s (Corona et al., 2006). A benefit of MPC is that it can utilise a cost function covering many variables and can satisfy constraints. This means that it can span a range of objectives. ACCs may wish to achieve good tracking performance, whilst reducing fuel consumption and keeping the driver comfortable.
Simulations of a heavy truck using such a Model Predictive Controller (Li et al., 2011) showed that it was possible to limit longitudinal acceleration to keep the driver comfortable, and by constraining the vehicle-following distance, fuel reductions were achieved as it could discourage other drivers to cut in and cause braking in the host vehicle - Figure 1-9. The car following model was based around a Generalised Vehicle Longitudinal Dynamics Model (following vehicle) and the interactions with a preceding vehicle. Desired acceleration 𝑎𝑓𝑑𝑒𝑠and inter-vehicle distances 𝑑𝑑𝑒𝑠are
determined by an upper level controller and the corresponding error signals are considered as variables in the cost functions to be optimised
1.4 Literature Review: Autonomous Speed Control 21
Figure 1-9 – MPC car following system (Li 2011)
Most of the reviewed literature fits into one of several categories of algorithm, as summarised by Table 3.
Table 3 – Speed Control Algorithms Summary
Algorithm Type Objective P, PI, PID Linear
Quadratic Dynamic Programming Model Predictive Car following - Machine Sivaji (2013) Shakouri (2011) Rajamani (2011) Shakouri (2011) Zhao (2011) Li (2011) Corona (2006)
Several approaches have already been investigated for determining a suitable and safe speed for a vehicle. Each has its own strengths and weaknesses. A feedback controller may use one of these algorithms for deciding a reference speed for feedback.
22 Introduction and Literature Review