Caregivers play a crucial role in assisting seniors having difficulty accomplishing activities of daily living (ADLs) due to physical or cognitive limitations. A global decline in the caregiver-to-senior ratio is making it increasingly more difficult to care for these seniors. Socially assistive robots are promising alternative technologies for supporting seniors in living independently. However, limited research has gone into developing a learning-based method for designing assistive robot behaviors. This thesis aims to:​ (1) identify the key features necessary for assistive robots supporting seniors with cognitive impairments in completing ADLs;​ and (2) develop a novel behavior-learning architecture to teach robots how to display assistive behaviors using expert demonstrations and personalize these learned behaviors to the seniorâ s cognition using reinforcement learning to increase task performance. Experiments with a socially assistive robot validated the robotâ s ability to learn and personalize new behaviors to a userâ s cognition from expert demonstration using the proposed architecture.