Multi-Agent Learning on Sensory data for Autonomous Systems

In many real world applications data is not static but arrives in data streams. On-line learning methods suits these applications by adjusting the learning policy incrementally as soon as new data arrives. In on-line settings agents never stop learning, while keep exploring the environment they learn, adapt and relearn in order to improve long term strategy. Constant exploration degrades system performance, so a balance is required between exploration and exploitation. This project aims to design a robust version of on-line reinforcement learning which overcomes the limitations of the sensory data such as missing values, outliers, and uncertainty etc. in order to improve convergence.