8. Experiments in Embedded Learning and Adaptation
8.3. Bringing it All Together
In this section the enhanced AFSM input associative mapping provided by the counterpropagation network in section 8.1 is combined with the output-adaptation mechanism of section 8.2. This is compared with a further possible solution to the problem of enhancing an AFSM in the form of a network-only solution, that of the real-valued output counterpropagation net as detailed in section 5 of chapter 7. Potentially both of these implementations provide a similar result in terms of enhancing the input-output mapping of an APSM, but the latter does not allow automatic tuning of output registers. Rather, it allows a varying output depending on input-
output mapping strengths and network weights. The experiments below will compare the merits of these two approaches.
Figure 8.24 shows the control structure of the BallPark and NearLin AFSMs for the new feedforward network enhanced system. Note that the N earLin AFSM subsumes BallParkand so utilises a counterpropagation network with a two-element output layer, one to generate the output value and the other to act as an output gating switch. Apart from this structural difference all network configuration parameters are the same as those used in the first (input only) implementation. N earL in B a iiP ark E H T O u tP u t 1 6 M e ta -P ix e is S c a n - L e v e l E H T V o lta g e L e v e l
Figure 8.24. BallPark and NearLin AFSM network control configuration.
Experiments
The two systems were each trained with the broken-surface sample and then tested with the same range of other surfaces (continuous, two-tone and broken two-tone) as used in section 8.1. Likewise a similar set of reduced input tests was performed by disabling the inputs to the networks in the same way as in section 8.1. Traces of the data relevant to system performance were made as before. Tests were run with the separate enhanced/adaptive AFSM processes and then with the complete feedforward-network AFSM process.
Continuous O utput AFSM Activity
Figure 8.25 shows a time series of output from the continuous-output network AFSMs after a leaming period on the broken-surface sample. The plots here show the EH TStepSize value output to the EH TO utputAFSM and illustrate that the rigid step-like output of the rule is superseded by a more flexible output. While the BallPark AFSM outputs generally match the basic mle output, it is interesting to note the expanded range of the N earLin AFSM actuation which was
additionally controlled by the learned output switch (figure 8.25). We deal with this in more detail in the discussion of results immediately below.
A NeaLin A^S M Output
W BOO 500 B: BdlPak A^SMCuput BOO -1000 10 1 151 2 0 Z -3000 -5000 5 -7000 ^
' R ule output EHT step size •AFSM rule state
Figure 8.25. Comparison of step sizes in EHT voltage as output by A: NearLin and B; BallPark. Basic rule output is compared with learned EHT step size. Plots relate to the broken-surface (surface sample
number 3).
Results
Figure 8.26 shows result graphs of the tests and compares previous tests with the two results (full feedforward network and enhanced input, adaptive output AFSMs) obtained here. It can be seen that the enhanced-input, adaptive-output configuration resulted in the higher reinforcement values for both the BallPark and NearLin AFSMs by a considerable margin. This was paid for by an accompanying small increase in the range of the reinforcement values. It is interesting to note the difference in AFSM output signals for these two configurations of controllers. It can be seen in figure 8.25 that the continuous-valued output AFSMs provided a much more gradual change in EHT step-size compared to the fixed-register-determined output of the enhanced/adaptive AFSMs. This is reflected in the relative sizes of standard deviations of EHT voltage output from the EHTOutput AFSM seen in figure 8.26. As a result the continuous-value output AFSMs appeared to have a much smoother control behaviour than that of the enhanced/adaptive type.
A: Mean R einforcement Values B : R an ge Values
No D cfo
BallPark N e a L in
Figure 8.26. Results of the Enhanced/Adaptive AFSM versus continuous-valued output AFSM tests compared with the best performers of the other tests in previous sections. As before, tests were conducted
on the broken-surface sample.
1 a E n h crcecV Adcptive Conrbnation Befüie L e u nil iQ _ 200 T 150 -- _ 1 0 0- - 50 0 100 0 1500 EHT V o ltc g e L w e l 2000
1 b. E nh cT ced/A dcptive Conrbinotion Afiei L ea n in g _ 20 0 T rr 100-- 500 1000 1500 EHT V o ltc g e L e / ë 2000
2o. Continuous O utputN etw ak AFSM Before L eaning
2b. Continuous Output Network AFSM After L ea n in g _ 200 j 150 -- — 100 - - 5 0 - - 0 2 0 0 - r 150 - r 100 - 50 - 0 5 00 10 0 0 1500 EHT V oltcgeL& Æ l 2000 5 00 1000 1500 EHT V o ltc g e L e / ë 2000
Figure 8.27. Scan-level / EHT-voltage state-space plots for the three test runs compared to the reactive- only system. 1. Enhanced/Adaptive AFSM, 2. Continuous-valued output AFSM. Plots relate to the
broken-surface sample (number 3).
It is therefore a matter of debate as to whether one technique should be chosen as a winner over the other, although comparison of the scan-level/EHT-voltage state-spaces in figure 8.27 provides another useful source of information. Here it would appear that the continuous-output- network AFSMs were significantly better. However, the answer to the question of the relative performances of these two techniques is most likely to be highly application-dependent. Some
other form of control activity would have to be implemented to address these issues, but that would be too large a programme to be within the scope of this thesis.