Currently, nearly all advanced automotive technologies operate under what are known as deterministic models. In a deterministic system, the outcome is dictated by a set of known initial conditions. These systems offer welcome predictability and repeatability. In contrast, probabilistic models yield different outcomes given the same initial condition, introducing an element of probability guided randomness. The latter model is considered necessary for autonomous vehicle functionality, but this is easier said than done.
Each model type has pros and cons. For starters, in the area of advanced automotive technologies, programming a deterministic model is more manageable. For example, automatic emergency braking systems are generally calibrated to recognize the rear of a preceding vehicle and apply braking when that vehicle slows and a crash is imminent. But how would that same system respond to a boulder rolling onto the roadway in the path of an equipped vehicle? The deterministic model is not well-suited for a curveball like this. The driver wants the same outcome – braking assistance – but the system only recognizes a certain set of predetermined conditions that do not include rolling boulders.
This simplistic example demonstrates one aspect of the limitations of deterministic modeling in the context of advanced automotive assistive systems. Probabilistic systems, on the other hand, may have the capacity to adapt and think more like humans. As of now, however, programming a probabilistic system to apply braking for new and random initial conditions, like a rolling boulder, is proving illusive. Human perception and decision-making (an example of a probabilistic system), although imperfect in many ways, is a highly sophisticated cognitive function that is difficult to simulate with technology.
We take our contextual understanding of the world for granted. People can easily distinguish boulders from cars, while computers recognize neither as what they really are. To a computer, both simply register as unique sensor data signals lacking any semantic meaning. As discussed above, applying a deterministic model, computers can be programmed to recognize certain patterns that dictate a defined outcome (e.g., braking for a slowing proceeding vehicle under ideal conditions). And when we see computers performing tasks normally attributed to humans in this way, it is easy to form expectations that far exceed the technology. After all, if your vehicle knows to stop short of rear-ending another vehicle, why wouldn’t it stop for all sudden hazards?
Another barrier to the introduction of advanced probabilistic models as the engines of advanced driving technologies is that they fly in the face of traditional automotive design strategy. Engineers strive for repeatable and predictable outcomes. A deterministic design can be calibrated, tested, and then re-calibrated as necessary to achieve the same desired outcome. This design-process paradigm works well for traditional mechanical systems and even for many advanced vehicle technologies. Testing, calibrating or certifying a probabilistic system, however, is inherently more challenging given the anticipated variability of outcomes. Variable outcomes, even if the model is statistically safer overall, may also prove difficult to defend in liability contexts.