Coaching: Learning and using environment and agent models for advice.

摘要

Coaching is a relationship where one agent provides advice to another about how to act. This thesis explores a range of problems faced by an automated coach agent in providing advice to one or more automated advice-receiving agents. The coach's job is to help the agents perform as well as possible in their environment. We identify and address a set of technical challenges: How can the coach learn and use models of the environment? How should advice be adapted to the peculiarities of the advice receivers? How can opponents be modeled, and how can those models be used? How should advice be represented to be effectively used by a team? This thesis serves both to define the coaching problem and explore solutions to the challenges posed.; This thesis is inspired by a simulated robot soccer environment with a coach agent who can provide advice to a team in a standard language. This author developed, in collaboration with others, this coach environment and standard language as the thesis progressed. The experiments in this thesis represent the largest known empirical study in the simulated robot soccer environment. A predator-prey domain and a moving maze environment are used for additional experimentation. All algorithms are implemented in at least one of these environments and empirical validation is performed.; In addition to the coach problem formulation and decompositions, the thesis makes several main technical contributions: (i) Several opponent model representations with associated learning algorithms, whose effectiveness in the robot soccer domain is demonstrated. (ii) A study of the effects and need for coach learning under various limitations of the advice receiver and communication bandwidth. (iii) The Multi-Agent Simple Temporal Network, a multi-agent plan representation which is refinement of a Simple Temporal Network, with an associated distributed plan execution algorithm. (iv) Algorithms for learning an abstract Markov Decision Process from external observations, a given state abstraction, and partial abstract action templates. The use of the learned MDP for advice is explored in various scenarios.