Bridging Classical Conditioning and Deep Reinforcement Learning: Advancements, Challenges, and Strategies for Autonomous Systems

Abstract

Many systems rely on Artificial Intelligence (AI), especially Deep Learning (DL), even though it is rarely used on its own to complete tasks. DL makes use of the Markov decision process as a framework for efficiently learning tasks. Theoretically, this procedure is similar to classical conditioning, which is how animals learn to connect actions and stimuli to goals. Deep Reinforcement Learning (DeepRL) was used in several studies to test DL skills in a variety of video games, showing that this technique can adapt to different tasks with little modification. However, those studies encountered major obstacles because of its large data requirements and expensive computational expenses, even if it was successful. Building on this, we examine the relationships between classical conditioning and DeepRL. Through careful manipulation of variables such as hyperparameters and maze designs, a robot was trained to navigate mazes as part of the experiment. DeepRL is not autonomous in this paradigm because the results showed that the Markov decision process and classical conditioning experience comparable challenges in tasks involving advanced planning and goal identification. The study also identifies the key areas that require improvement, highlighting the shortcomings of existing AI systems and offering strategies for boosting their autonomy.