Neural network agents trained by declarative programming tutors

This paper presents an experimental study on the development of a neural network-based agent, trained using data generated using declarative programming. The focus of the study is the application of various agents to solve the classic logic task - The Wumpus World. The paper evaluates the effectiveness of neural-based agents across different map configurations, offering a comparative analysis to underline the strengths and limitations of these approaches. We discuss the quantitative and qualitative aspects of these agents in scenarios that require generalization. For a concise comparison, we present the performance and resource utilization of different agents as follows: The Prolog-based agent showed a base task win rate of 61 %, which dropped to 5 % in a modified task setting, requiring 13KB of memory. The Q- Learning agent achieved a 2 % win rate in the base task, with the modified task performance not applicable, and a memory requirement of 67KB. An agent based on a Convolutional Neural Network (CNN) recorded a 44% win rate on the base task and 32% on the modified task, consuming 134KB of memory. The Deep Q-Network (DQN) agent displayed a 56% win rate in the base task and 46 % in the modified task, necessitating a substantial amount of memory, 284MB.