Learning from Demonstrations and Human Evaluative Feedbacks: Handling Sparsity and Imperfection Using Inverse Reinforcement Learning Approach

Table 2

Setting used to study different aspects of our framework () in the simulated navigation domain.

Aspect	Demonstration		Feedback error	Learner policy type
	Step number	Optimality degree
Comparison (5.1.1)	100	60% (point A5)	0	Probabilistic
	20	100% (point B2)	0	Probabilistic
Nonoptimal demonstration effect (5.1.2)	100	Different (points: A1–A8, B10, C)	0	Probabilistic
Sparse demonstration effect (5.1.3)	Different (points: B1–B10, C)	100%	0	Probabilistic
Learn only from feedback (5.1.4)	No demo (point C)	—	0	Probabilistic
Effect of the policy execution (5.1.5)	100	60% (point A5)	0	Probabilistic, greedy, random
Effect of feedback error (5.1.6)	100	60% (point A5)	ε = 0, 0.1, 0.5	Probabilistic

Note: learning from the collected feedbacks is done in the batch learning mode.