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

Table 4

Results of the learned navigation experiment by the E-puck robot in two different environments, with interaction steps with the environment before the learner policy is updated. Here, in experiments 1 and 3, demonstrations and feedbacks are provided in the same environment, and then, the performance of the learned reward function is examined in the second environment. Experiment 2 is done by providing sparse demonstrations in one environment and feedbacks in another.

Experiment	Demonstration		Environment 1				Environment 2
			Number of feedbacks		EV of	EV of teacher policy	Number of feedbacks		EV of	EV of teacher policy
	Type	Number of	All feedbacks	Negative feedbacks			All feedbacks	Negative feedbacks
1	Abundant, nonoptimal	100	135	51	2.735	2.824	—	—	3.145	3.279
2	Sparse, near-optimal	22	—	—	2.749	2.824	239	70	3.157	3.279
3	No demo	—	311	181	2.761	2.824	—	—	3.188	3.279