Skip to content

DA1. Self Driving Car Model

Statement

Can computers learn and adapt?

Given a scenario where a computer is installed in a car and the goal is to become a self-driving car. You could code/write all the possible rules or you could provide basic rules and have the computer/car learn from experience. Discuss, how does machine learning allows a computer to learn and perform a task without explicit programming? Which method is more efficient and why?

Answer

Introduction

Computers can learn. Machine Learning (ML) is a branch of Artificial Intelligence (AI) that is responsible for teaching agents (in this case, an instance of a Self-Driving Car SDC) to learn and adapt without being explicitly programmed based on concepts such as neural networks, decision trees, logistic regression, and gradient-boosted trees.

In this text we will explain the dimensions of the agent design space, then we will discuss those dimensions in the context of a Self-Driving Car SDC model, and finally, we will conclude with answering the questions in the prompt.

Agent Design Space

These are some of the dimensions in the design space of agents, according to Poole and Mackworth (2017):

  • Modularity: separating agent functionality into modules (or smaller agents) that can have flat, hierarchical, or modular structures.
  • Planning horizon: the time span over which the agent plans its actions. It can finite, infinite,, indefinite, or none. Each module in a modular agent can have its planning horizon.
  • Representation: the way the agent represents the world, which can be using states and features (previously recorded states), or using relational logic (desired relations between individual and their properties).
  • Computational limits: the computational resources constraints on the agent’s rationality that makes it perfect (if there are no limits) or bounded (if there are limits).
  • Learning: the agent’s ability to learn from experience; there are given knowledge that should be recorded before the agent starts, and others that agent need to extract from the environment in real-time.
  • Sensing Uncertainty: the uncertainty in the agent’s observations of the world which can be fully or partially observable.
  • Effect Uncertainty: the uncertainty about the state of the world after the agent’s actions, which can be deterministic (only depends on the action and the prior state) or stochastic (depends on the action, the prior state, and some random factors).
  • Preferences: the agent’s sorting functionality of the possible actions that it should or can take which can be based on goals, ordinal, or cardinal preferences.
  • Number of agents: the number of agents in the environment and how the agent interacts with them.

Self-Driving Car (SDC) Model

Let’s now discuss the dimensions of the agent design space in the context of a self-driving car (SDC) model:

  • Modularity: SDC is a complex model that needs smaller modules for functionality such as localization, path planning, collision avoidance, internal monitoring, and communications with other cars and entities on the road. All of these modules need to be integrated under a central control system that makes decisions on normal and emergency situations.
  • Planning horizon: SDC needs to plan its actions in real-time based on the current state of the road and the car. Some modules need infinite planning horizon (e.g., collision avoidance that needs to keep working all the time) while others need finite planning horizon for different periods of time (e.g., path planning spans over the journey time).
  • Representation: SDC needs to represent the world using relations as the world state is dynamic and too complex to be loaded on a its belief system. Relations are functions that embed reasoning about the world, so instead of storing all the locations of traffic lights, the SDC can store a relation that tells it when it sees a traffic light, it should evaluate its color and act accordingly.
  • Computational limits: SDC is bounded by computational limits as it needs to make decisions in real-time; when there is no time to compute the optimal action, the SDC needs to make the best decision that it finds now to keep the car and the passengers safe, and record the situation for future learning.
  • Learning: SDC needs some given knowledge about road rules, emergency procedures, and car mechanics. There are other knowledge that the SDC needs to learn from experience such as other car models available on the road, if the law changes, and specific road conditions.
  • Sensing Uncertainty: SDC needs to deal with uncertainty in its sensors as they can be affected by weather conditions, communication problems between different parts of the car, and other cars on the road. The SDC needs to understand its partially observable world and take all inputs with a probability distribution.
  • Effect Uncertainty: SDC needs to deal with uncertainty in the effect of its actions as the world is stochastic and the SDC can’t predict all the possible outcomes of its actions. For example, signaling to change lanes where the other car doesn’t respond, in this case, the SDC should not take the signaling as deterministic action.
  • Preferences: SDC needs to sort its actions based on its goals of keeping itself and passengers safe, following the road rules, and reaching the destination in the shortest time.
  • Number of agents: SDC needs to interact with other road users and their intentions which may not be always cooperative.

Conclusion

To conclude, yes, computers can learn and adapt. Machine Learning (ML) allows a computer to learn and perform a task without explicit programming as it is unachievable with a complex model such as a Self-Driving Car (SDC) to code all the possible rules.

Instead, the world is represented into relations where the SDC identifies the individuals and their properties and intentions in real-time and make decisions while keeping track of the outcomes which is then stored and used as feedback in the learning process.

References

  • Poole, D. L., & Mackworth, A. K. (2017). Artificial Intelligence: Foundations of computational agents. Cambridge University Press. Chapter 1 - Artificial Intelligence and Agents. https://artint.info/3e/html/ArtInt3e.Ch1.html