Extraction of information from ai design decisions - theoretical and practical case study

Abstract

In recent years, the emergence of Artificial Intelligence (AI) has revolutionized numerous fields [1], such as the healthcare and automotive industries, and the technologies that use AI have become part of our daily routines. This surge in AI use in different fields has raised many issues because of a lack of transparency regarding the black-box nature of the AI algorithms and, specifically, how the neurons of the neural networks in the AI algorithms interact to conclude outcomes [2]. This problem is especially notable in the automotive industry, where AI vehicles increasingly integrate AI into their systems, especially the Advanced Driver Assistance Systems (ADAS), to improve safety and driver convenience. For engineers, consumers, and regulators, its essential to understand the reasoning behind AI decision-making for many reasons. Accountability, among many other reasons, is considered significant in this context; for example, who will be responsible for accidents when they happen? Is it the person behind the steering wheel or the car maker? [3]. Because of those reasons, the significance of revealing the AI decision-making process by extracting information from AIs decision-making algorithms cannot be overstated, especially in the context of ADAS and, more specifically, in Rear-End Collision Avoidance (RECA) systems. This thesis is broken down into several sections: 1. We trained a reinforcement learning agent to operate a Rear-End Collision Avoidance (RECA) System using the Reinforcement Learning toolbox of MATLAB [4]. We used the code and Simulink model of the example of the Collision Avoidance System (Train DDPG Agent for Adaptive Cruise Control) in the Reinforcement Learning toolbox in MATLAB [5] and the car dynamics of the example of the Adaptive Cruise Control System Using Model Predictive Control in the Model Predictive Control Toolbox in MATLAB [6]. Then, we modified that examples code and Simulink model to fit my application. 2. We developed MATLAB code to extract the trained agent from the Reinforcement Learning (RL) algorithm. 3. We verified the extracted RL agent by simulating it using Simulink and then without the RL agent to confirm its necessity for the RECA system. 4. We formulated the mathematical relationship between the trained agents inputs and output by fitting it to a Linear Regression Model (LRM) using a MATLAB function called fitlm [7]. 5. We verified the LRM by simulating it using Simulink. 6. We used two scaled robotic cars to verify the LRM to confirm the reliability of the extracted information. 7. We analyzed the LRM using the State Space Control technique to determine the basis function and better understand the foundation control law of the RL algorithm. 8. In the last part, we verified the basis functions formula by simulating it using Simulink and then simulating it by excluding some of the basis functions as an extra step to confirm the necessity of all the basis functions in the set. This in-depth theoretical and practical analysis is a significant step towards unraveling the ambiguity of AI algorithms in different fields, especially in the automotive field, by shedding light on the black box of AI algorithms and complex decision-making processes of AI.