Basics of Autonomous Vehicle Technology and Energy Consumption Ramifications
1.0 Introduction
Autonomous (driverless) vehicle technology (“AV”) is developing at a rapid pace. Waymo has developed software and hardware sensors to detect pedestrians, cyclists, road works, and vehicles (Waymo). Ride-hailing company Uber is committed to buying 24,000 Volvo SUVs which it will equip with AV driving technology (Frangoul). Indeed, self-driving cars are already being tested on the road: GM Cruise vehicles were on California roads (and involved in 13 crashes) in 2017 (Shepardson).
Although safety issues with AV are in the forefront, another significant question is the impact their widespread usage will have (if any) on energy consumption. My research suggests that the answer latter question is currently indeterminate. In a 2013 report (see Appendix A), the United States Department of Energy stated that the possible energy impact of AV ranges from 87% fuel savings to a 217% increase in energy usage (National Renewable Energy Laboratory). (Hereafter I shall refer to this report as “NREL.”)
In this paper, I will explicate two distinct approaches to computing, viz., computational and neural, and show how each is important to autonomous vehicle development. I will then discuss several features of autonomous vehicle systems and what impact each might have on energy consumption. Finally, I will show why the total effect on energy consumption is unknown at this point.
2.0 Two Distinct Approaches to Computing
Because AVs are designed to be driver-less, or at least not require driver intervention to move safely and efficiently from place to place, we might conclude that the computers which operate them must have human-like capabilities. Let’s examine two models for computing in search of a human-like model.
2.1 Computational
Traditional computational computing involves serial processing of if-then statements (see fig. 1). Although parallel threads can be used, each thread travels a definitive path from input to output.
Further, computational systems manipulate only syntactic properties, i.e., the shape of symbols. The system doesn’t understand meaning, i.e., the semantic properties of symbols. However, syntactic processing respects semantics. For example, a computer program can begin with two symbols:
A. All philosophers are funky.
B. Fodor is a philosopher.
And arrive at this conclusion:
C. Fodor is funky. (Ravenscraft 83–84)
The program, operating only on syntax, manipulated symbols alone (in a deductive argument) to arrive at a valid conclusion (output) with semantic meaning. Because of this ability of computational systems, some have argued for the Computational Theory of Mind which holds that the human brain itself is nothing more than a computational system (Ravenscraft 81–96). However, others such as John Searle, have famously argued that human minds must be more than computational systems, because consciousness and meaning exist and cannot be accounted for in computational systems which only manipulate symbols[1].
Human brain-like or not, AVs indeed rely on computational computing for some of their functions, as discussed below.
2.2 Neural Network
Also known to philosophers as connectionism, neural network theory inserts a probabilistic (as distinct from deterministic) hidden layer between the input and output (see fig. 2). Additionally, neural network threads run in parallel; for these reasons neural networks tend to accomplish tasks more quickly than computational systems. Connections between inputs and the hidden layer are configured with connection weights, which are constantly re-weighted in an ongoing learning process designed to produce the most accurate output[2]. For example, a neural network can distinguish between mine echoes and rock echoes (Ravenscraft 98).
Distinguishing between objects with high specificity, as well as self-learning, are features that AVs require, which is why neural networks are employed in some AV functionality, as described below. Further, neural networks can generalize from previous cases to similar new cases and can process distributed tasks locally (Ravenscraft 101–103). These additional features of neural networks are similar to natural human capabilities, are thus are also useful in AV systems, as we shall see[3].
Regarding the distinction between static data-driven computational systems and learning-capable neural networks as applied to autonomous vehicle systems, one hobbyist website states succinctly:
. . . we should avoid conceptualizing self-driving vehicles as machines which are controlled by a detailed, exactly specified and in principle comprehensible software program. Instead we should conceptualize their behavior as being the result of a long and varied program of learning. The capability of such cars can be analyzed through simulation and testing but not just by examining its source code. (Hars)
3.0 System Communications
In an autonomous vehicle system of systems, we are concerned with communications from one car to other cars (“C2C”), cars to the central infrastructure (“C2I”), and infrastructure to cars (“I2C”) (Uhler and Uhl Sec. 10). An example of C2C is negotiation between two or more cars to facilitate safe navigation. An example of a simple C2I communication is a “Can I go?” query; the answer is returned I2C.
C2C is facilitated by the distributed nature of a neural network. Some of the C2I and I2C functions, as we shall see below, are in theory better suited to a deterministic computational system. However, due to capacity limits in processing resources and speed, neural networks may be more practical even where computational systems are superior in theory. Further, local decisions by a particular automobile, e.g., obstacle avoidance, are best suited to distributed neural networks because of their capability of local processing.
4.0 Assessing Energy Impact
To assist in assessing potential effects of autonomous vehicle systems on energy consumption, I utilized NREL. The report identifies various features and phenomena associated with AVs and their operation and makes educated guesses as to the potential energy impact of each feature. It categorizes factors as either use intensity (“UI”), energy intensity (“EI”), or fuel intensity (“FI”). It further distinguishes between AVs under private versus shared ownership, but I shall combine those scenarios for the purposes of this paper. Finally, NREL uses an equation to predict fuel demand effects (see fig. 3) dependent upon which of various scenarios come to fruition (see fig. 4).
4.1 Effects That May Reduce Energy Consumption
4.1.1 Efficient routing
According to NREL, efficient routing and traffic avoidance can reduce EI 20%. Efficient routing can be accomplished in two ways (Caditz):
4.1.1.1 Computational method
In theory, if/then computation is the best method because given a set of conditions (e.g., start point, end point, competing traffic) the most efficient route can be definitively determined. Further, a central command point is required for C2I and I2C communications. However, such computations are time and resource intensive and as such, practical limits may be reached. It therefore may be necessary to employ a faster and probabilistic neural network.
An additional factor in whether computing should be performed on infrastructure level versus distributed to car level relates to the philosophical goal of the AV system: Is it to optimize the total utility of the system, or is it be left to each car to optimize its own utility (Caditz)? The former approach might be analogized to an egalitarian political system; the latter to a competitive, libertarian system. How this philosophical debate is resolved speaks to the type of computer technology required in system implementation, and whether I2C or C2C communications takes priority.
4.1.1.2 Neural network method
The learning-capable neural network facilitates an evolving database of situations encountered (e.g., how many cars are on a route or in an intersection, how much congestion occurred, minimal total travel time) and can rank each situation. The car can then constantly choose the highest ranked solution given the current conditions, reducing the dependence on central command and the needs for C2I and I2C communications.
4.1.2 Obstacle and collision avoidance
NREL does not assess energy impact of this feature of AVs. However, to the extent collisions are increased or reduced compared to current levels, there should be a corresponding increase or reduction in energy consumption associated with vehicle waste disposal, repairs, replacement vehicle production, and medical care.
4.1.2.1 Computing method
A pre-configured database of all possible obstacles and appropriate reactions is impractical, leaving the learning neural network as the best option. Further, a neural network allows for local decisions (including C2C) as to when braking or other avoidance measures are required (Caditz).
4.1.3 Efficient driving
This includes smoother starts and stops, as well some stop elimination. According to NREL, this effect may reduce EI by 20–30%.
4.1.4 Platooning
According to NREL, close following at high speed reduces drag and may reduce EI by 10%.
4.1.5 Lighter vehicles
According to NREL “Very few crashes could enable very light vehicles for many duty cycles,” and reduce EI by 45%.
4.1.6 Higher vehicle occupancy
Automated carpooling could facilitate more people per vehicle, decreasing UI 12–020% according to NREL.
4.1.7 Reduced parking searching
According to NREL, fewer vehicles and routed self-parking could reduce UI by 4% and cut wasted fuel in half.
4.1.8 Electrification
AV is not necessary for electrification of vehicles. However, NREL asserts that AV solves a problem of electric vehicles, viz., geographical range limits: The AV infrastructure could deploy shared vehicles to match the users need and the range limit. Presumably, multiple fully-charged vehicles could be deployed for long trips. NREL suggests a possible 75% reduction in FI as a result of electrification of 75% of total vehicles.
4.2 Effects That May Increase Energy Consumption
4.2.1 Increased travel
This potential effect of AV threatens to offset many of of the energy savings effects, according to NREL. At least two reasons for increased travel can be identified:
4.2.1.1 Accessibility to underserved populations
Shared vehicles may make AV travel accessible to those who cannot afford vehicle purchase. Further, AVs may not be prohibitive to those previously unable to drive, e.g., youth, elderly, and disabled. The result could be a 70% UI increase.
4.2.1.2 Increased discretionary travel
Faster travel times and more pleasant travel could increase utilization and allow for extended practical travel and commute distances, potentially resulting in a 50% UI increase.
4.2.2 Faster driving
Improved routing and safety could lead to faster speeds and reduced energy efficiency. This could increase EI by 30% according to NREL.
4.2.3 Energy cost of computing resources
There will be energy consumption by the AV computing infrastructure[4]. This cost could be substantial: For example, blockchain technology, a computing infrastructure which underlies bitcoin, is famously energy-intensive, wasteful, and is said to “eat up more energy than Argentina” in a year (Zhao).
5.0 Summary
We examined two computing models — computational and neural network — each of which has its place in AV features which will likely impact energy consumption. We also explored the types of communication required between entities: C2C, C2I, and I2C; and how which takes priority is linked to the underlying philosophy of the AV system. But the consequence of AV on energy consumption is the subject of widely varying predictions: NREL predictions are dramatically different depending on which of the multiple scenarios is ultimately adopted (see figure 4). If all possible fuel demand-decreasing effects come to fruition, fuel demand would decrease 87%; but if all possible fuel demand-increasing effects come to fruition fuel demand would increase 217%. NREL concludes that the wide range of potential outcomes “emphasizes the importance of considering energy impacts in AV deployment strategy.” Further, as noted above, NREL does not anticipate all possible energy impacts of AV, including the energy cost of AV computing infrastructure itself — which could be substantial.
Appendix A: Department of Energy Report
Works Cited
Caditz, David M. Autonomous Vehicle Routing Researcher Michael Robert Caditz. 27 April 2018. Skype Interview.
Frangoul, Anmar. “Autonomous vehicles will transform the way driving is regulated: Here’s how.” 10 Jan 2018. CNBC. Web. 29 April 2018.
Hars, Alexander. “Top misconceptions of autonomous cars and self-driving vehicles.” 29 July 2017. Driverless car market watch. Web. 28 April 2018. <http://www.driverless-future.com/?page_id=774#programming-model>.
National Renewable Energy Laboratory. Workshop on Road Vehicle Automation. Stanford: U.S. Department of Energy, 2013. Print. <https://www.nrel.gov/docs/fy13osti/59210.pdf>.
Ravenscraft, Ian. Philosophy of Mind: A Beginners Guide. New York: Oxford University Press, 2005.
Scientific American. “What is a neural network and how does its operation differ from that of a digital computer? (In other words, is the brain like a computer?).” n.d. Scientific American. Web. 28 April 2018. <https://www.scientificamerican.com/article/experts-neural-networks-like-brain/>.
Shepardson, David. “GM’s self-driving cars involved in six accidents in September.” 4 Oct 2017. Reuters. Web. 29 April 2018.
Uhler, Werner and Heribert Uhl. Driver assistance method and apparatus. Patent WO2010000521A1. 04 May 2009. Web. <https://patents.google.com/patent/WO2010000521A1/en>.
Waymo. n.d. Web. 29 April 2018. <https://waymo.com/>.
Zhao, Helen. “Bitcoin and blockchain consume an exorbitant amount of energy. These engineers are trying to change that.” 27 Feb 2018. CNBC. Web. 29 April 2018. <https://www.cnbc.com/2018/02/23/bitcoin-blockchain-consumes-a-lot-of-energy-engineers-changing-that.html>.
[1] See Searle’s Chinese Room thought experiment (Ravenscraft 91–94).
[2] For a detailed description of neural network weighting, see Ravenscraft (97–100).
[3] We should not overstate our claim that neural networks are similar to human brains. According to one article in Scientific American, “These networks are ‘neural’ in the sense that they may have been inspired by the brain and neuroscience, but not necessarily because they are faithful models of biological, neural or cognitive phenomena” (Scientific American).
[4] Infrastructure energy cost is not mentioned in NREL.
