KNACTOR: architecture for a learning intelligent agent

KNACTOR (know+actor) is an intelligent agent prototype that learns how to represent and control the state of its dynamic environment. Beginning only with knowledge of its available sensor and actuator values, KNACTOR explores its environment and creates a multilevel environment model consisting of multiple coordinated submodels. The model is developed on a blackboard and is then used to plan actions leading to a goal state which KNACTOR learns to associate with a reward signal applied by the experimenter. Both probabilistic and deterministic techniques are employed, and some of the heuristics have a genetic learning flavor. Multiple models enable the agent to benefit from efficient processing within small state spaces, taking advantage of the conveniences and favorable features of the available representations that are appropriate in different circumstances.<
>