Seth O. Rogers
We present an approach which enables an agent to learn to achieve goals in continuous environments using a symbolic architecture. Symbolic processing has an advantage over numerical regression techniques because it can interface more easily with other symbolic systems, such as systems for natural language and planning. Our approach is to endow an agent with qualitative "seed" knowledge and allow it to experiment in its environment. Continuous environments consist of a set of quantitative state variables which may vary over time. The agent represents goals as a user-specified desired value for a variable and a deadline for its achievement. To determine the correct action given the current situation and goals, the agent maps the numbers to symbolic regions, then maps these regions to an action. The learning task of the agent is to develop these mappings.