Electronic Theses and Dissertations



Document Type


Degree Name

Doctor of Philosophy



Committee Chair

Bonny Banerjee

Committee Member

Dale Bowman

Committee Member

Madhusudhanan Balasubramanian

Committee Member

Eddie Jacobs


Predicting the state of an agent's partially-observable environment is a problem of interest in many domains. Typically in the real world, the environment consists of multiple agents, not necessarily working towards a common goal. Though the goal and sensory observation for each agent is unique, one agent might have acquired some knowledge that may benefit the other. In essence, the knowledge base regarding the environment is distributed among the agents. An agent can sample this distributed knowledge base by communicating with other agents. Since an agent is not storing the entire knowledge base, its model can be small and its inference can be efficient and fault-tolerant. However, the agent needs to learn -- when, with whom and what -- to communicate (in general interact) under different situations.This dissertation presents an agent model that actively and selectively communicates with other agents to predict the state of its environment efficiently. Communication is a challenge when the internal models of other agents is unknown and unobservable. The proposed agent learns communication policies as mappings from its belief state to when, with whom and what to communicate. The policies are learned using predictive coding in an online manner, without any reinforcement. The proposed agent model is evaluated on widely-studied applications, such as human activity recognition from multimodal, multisource and heterogeneous sensor data, and transferring knowledge across sensor networks. In the applications, either each sensor or each sensor network is assumed to be monitored by an agent. The recognition accuracy on benchmark datasets is comparable to the state-of-the-art, even though our model has significantly fewer parameters and infers the state in a localized manner. The learned policy reduces number of communications. The agent is tolerant to communication failures and can recognize the reliability of each agent from its communication messages. To the best of our knowledge, this is the first work on learning communication policies by an agent for predicting the state of its environment.


Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest