Building an Adaptive Learning Agent in V-world

摘要

This paper describes Learner, an agent that learns from its experiences. Initially Learner knows neither the effects of its actions nor the properties of the objects in its world. Furthermore, Learner does not have an explicit representation of its environment. Instead it uses the world as its own model. The agent is given restricted abilities to perceive and to move, and the rules of death and survival. It explores its world, and through trial-and-error learns its properties. As a result Learner can survive indefinitely in low-risk worlds and lives long enough to be called intelligent in the highest-risk worlds. Learner is a reinforcement learning agent: it does not have a teacher; instead it employs a condition-action component, where the action evaluation is the reinforcement. This is an LPA Prolog implementation developed in V-world, an agent-testing computer environment featuring various simulated worlds.