We present an approach to the problem of learning by observation in spatially-situated tasks, whereby an agent learns to imitate the behaviour of an observed expert, with no direct interaction and limited observations. The form of knowledge representation used for these observations is crucial, and we apply Qualitative Spatial-Relational representations to compress continuous, metric state-spaces into symbolic states to maximise the generalisability of learned models and minimise knowledge engineering. Our system self-configures these representations of the world to discover configurations of features most relevant to the task, and thus build good predictive models. We then show how these models can be employed by situated agents to control their behaviour, closing the loop from observation to practical implementation. We evaluate our approach in the simulated RoboCup Soccer domain and the Real-Time Strategy game Starcraft, and successfully demonstrate how a system using our approach closely mimics the behaviour of both synthetic (AI controlled) players, and also human-controlled players through observation. We further evaluate our work in Reinforcement Learning tasks in these domains, and show that our approach improves the speed at which such models can be learned.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:699152 |
| Date | January 2016 |
| Creators | Young, Jay |
| Publisher | University of Birmingham |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://etheses.bham.ac.uk//id/eprint/7096/ |
Page generated in 0.013 seconds