Interactive Imitation Learning extends standard imitation-learning by allowing the learner to query the expert interactively during training, rather than learning only from a fixed dataset of demonstrations. This addresses the exposure-bias problem by ensuring the learner receives guidance on states it actually encounters during rollout, similar to how on-po...
- Related
- References
Interactive Imitation Learning
Interactive Imitation Learning extends standard imitation-learning by allowing the learner to query the expert interactively during training, rather than learning only from a fixed dataset of demonstrations. This addresses the exposure-bias problem by ensuring the learner receives guidance on states it actually encounters during rollout, similar to how on-policy methods in distillation address distribution mismatch.
Related
- exposure-bias — Interactive IL directly addresses exposure bias by querying experts on learner-visited states
- imitation-learning — Broader imitation learning paradigm
- on-policy-distillation — OPD similarly addresses distribution mismatch through on-policy rollouts