Lifted/Relational Reinforcement Learning

The high level overview of this project is to build relational models for sequential decision making. The main bottleneck of approximating utility functions in decision making is that it requires sampling large number of states to reasonably approximate the function. However, most real world domains have large state space and are relational in nature. Learning rich relational representations in such domains can help in capturing useful interactions between entities and similar states can be clustered together to learn lifted relational models which can be useful and efficient utility function approximators in large state space. As part of this project, we focus on efficient relational representation learning and relational sequential decision making for hybrid domains. We also wish to combine structured symbolic representations with deep models to create powerful, efficient and interpretable deep utility function approximators. We are also working on knowledge augmented Imitation learning where the goal is to learn the demonstrator’s noisy policy using advice from multiple experts.