Energy savings

In Table 3, we give energy savings (ES) against the baseline. These include the simple, lazy, perfect, psychological, and history techniques. In the history technique, the pessimism factor $\alpha$ is assumed to be $0.4$ and the last seven delay records are used starting with the psychological predictions. The average underestimation error (AUE) (the fraction by which the user delay is underestimated), percentage of lazy error (PLE) (the percentage of all predictions that are over-estimations) and percentage of serious lazy error (PSLE) are also reported for the psychological and history techniques. We adopt $50ms$ as the human perceptual threshold.

Table 3: Energy savings based on predicted user delays

Benchmark	User	Simple	Lazy	Perfect	Psychological				History(0.4)
		ES(%)	ES(%)	ES(%)	ES(%)	PLE(%)	PSLE(%)	AUE(%)	ES(%)	PLE(%)	PSLE(%)	AUE(%)
Calculator	1	17.9	18.8	35.1	27.3	0.0	0.0	28.4	20.1	0.0	0.0	64.5
	2	17.9	15.3	34.3	22.9	0.0	0.0	39.9	19.3	0.0	0.0	60.5
Filebrowser	1	17.6	32.9	37.7	23.7	0.0	0.0	51.3	21.8	2.7	2.7	64.8
	2	17.7	34.4	38.2	21.0	0.0	0.0	61.9	19.4	0.0	0.0	73.9
Go	1	17.5	36.0	38.2	18.9	0.5	0.0	77.5	21.4	10.6	9.7	66.5
	2	15.8	25.3	32.9	18.9	2.3	1.2	59.9	19.0	1.7	0.0	63.2
Solitaire	1	18.2	35.2	39.3	22.8	3.7	2.9	61.5	22.8	7.9	6.2	64.8
	2	17.9	32.4	38.0	19.6	8.7	5.5	68.9	21.4	14.2	12.0	65.7

Notably, Lazy's energy savings are closest to Perfect's in three of the four benchmarks. Only in Calculator, it is less energy-efficient than the other techniques due to the fact that user delays for Calculator tend to be much shorter and predictable (see Section VI). However, its PSLE is almost 100%. Therefore, Lazy will be useful only when system delay is tolerable and user delay is relative long, as in the case of map viewer and text reader applications.

For the psychological technique, the PLE is very low since this model is pessimistic. However, for the history technique, the pessimism factor controls the tradeoff between the PLE and ES. Fig. 4 shows how the ES changes with PLE for Calculator and Filebrowser as the pessimism factor varies from $0.2$ to $1.3$ for the history technique (His). It also shows the tradeoff point for the psychological technique (Psy). To achieve the same energy saving as the psychological technique, the history technique needs to make a lot more lazy errors. This demonstrates the superiority of the psychological technique when avoiding lazy errors is important.

Figure 4: Tradeoff between percentage of lazy errors and energy savings for Calculator and Filebrowser.

To show the benefit of using user interface information in the history technique, Fig. 4 also shows the tradeoff curve (OldHis) for the conventional history-based idle time prediction technique, which uses the last seven observed delays to predict the next delay without regard to STD states. It is clear that STD state-aware delay prediction is better than conventional prediction. Moreover, the Calculator and Filebrowser's user interfaces are relatively simple. We expect the advantage of using state-aware delay prediction to be more for more complicated user interfaces.

It is worth noting that the psychological technique performs better for Calculator and Filebrowser than Go and Solitaire compared to the history technique in terms of ES, AUE, and PLE/PSLE. This is due to the fact that operating Calculator and Filebrowser is cognitively much simpler and their cognitive processes are better modeled by the Hick-Hyman Law.