Adaptation from Sparse Observations through Detection and Correction of Model Anomalies

To make intelligent decisions, robots often use models of the effects of their actions on the world. Unfortunately, it is often infeasible to have the perfect knowledge and computational resources required to create globally accurate models. This may lead to divergence between planned and actual execution, often resulting in suboptimal task performance. We propose an execution monitoring approach that enables robots to detect unmodeled modes of the system and correct their models accordingly. In particular, we address the problem of finding and adapting to regions of a state-action feature space in which outcomes of actions observed during execution deviate from the stochastic expectations used to select those actions. We evaluate our approach in the adversarial domain of autonomous robot soccer. Categories and Subject Descriptors I.2.9 [Robotics] Keywords Multi-robot systems; Robot planning and plan execution; Single and multi-agent learning; Fault tolerance and resilience.

• Type: Journal
• Year: 2015
• Where: ATAL
• Authors: Juan Pablo Mendoza
• Keywords: ATAL 2015, Intelligent Agents