The addition of the new rule into the results doesn’t appear to have done much, save slowing the results down. Of course, this could be a result of something else, but nonetheless…
Anyway, I have already run an experiment that I neglected to comment on regarding the variance. The results have far too much variance to be able to accurately represent a technique in a single run, so all forthcoming experiments will be averaged over ten separate runs. The performance will accommodate the average, minimum and maximum scores (min and max defined by their total sum over the episodes, rather than min or max per episode).
I believe I have fully decoupled the GUI interface, so it should (theoretically) run faster, although with an average of 10 runs, it’ll still take quite some time. Hopefully the 64-bit machines will be fast.
The experiments to run now are:
Hand-coded
Random rule (from a given rulebase)
Random rule with constant integration
Random rule with individual regeneration
The results for the current experiments are:
Hand-coded, 50 episodes:
24697.533
29686.203
38102.8
41015.8
42552.137
48541.395
51106.195
53647.66
56438.066
63126.926
65339.465
67878.0
68647.59
67974.53
69542.94
70399.45
71693.4
72586.06
72353.59
73928.8
73319.0
73373.87
74931.13
75117.41
78882.92
81736.81
82270.0
85026.34
83788.27
84964.8
85803.59
86235.74
87341.73
85669.6
86840.46
87001.06
87746.86
88220.99
86439.79
87639.75
86317.46
86380.01
87833.87
86505.2
86938.66
87895.26
87307.66
87009.4
87999.0
87868.0
Again:
22198.936
32501.87
37092.6
42389.33
46743.27
48682.594
50626.477
54815.99
56553.8
62885.6
65058.797
67507.67
69595.47
69864.52
73136.07
72886.2
74585.54
73505.53
75439.94
77619.33
80422.53
78542.59
81496.68
84225.54
84251.74
84112.73
83835.6
86120.07
85196.39
84760.13
86625.14
86181.86
87454.26
86297.81
87055.81
87148.33
88829.35
89653.39
87957.26
87855.94
89109.34
90371.15
89762.19
89602.47
88323.8
88718.66
90218.33
89422.61
88796.32
89956.08
One more (to come):