Python 3, 2·267 +たす 510193 =わ 510727

Predictor

def p():
 d={};s=b''
 while 1:
 p={0:1};r=range(len(s)+1)
 for i in r:
 for c,n in d.setdefault(s[:i],{}).items():p[c]=p.get(c,1)*n**b'1円6円\f36円AcWuvY_v`270円~333円~'[i]
 c=yield max(sorted(p),key=p.get)
 for i in r:e=d[s[:i]];e[c]=e.get(c,1)+1
 s=b'%c'%c+s[:15]

This uses a weighted Bayesian combination of the order 0, ..., 16 Markov models, with weights [1, 6, 12, 30, 65, 99, 87, 117, 118, 89, 95, 118, 96, 184, 126, 219, 126].

Test code

with open('whale2.txt', 'rb') as f:
 g = p()
 wrong = 0
 a = next(g)
 for b in f.read():
 wrong += a != b
 a = g.send(b)
 print(wrong)

Python 3, 2·267 +たす 510193 =わ 510727

Predictor

def p():
 d={};s=b''
 while 1:
 p={0:1};r=range(len(s)+1)
 for i in r:
 for c,n in d.setdefault(s[:i],{}).items():p[c]=p.get(c,1)*n**b'1円6円\f36円AcWuvY_v`270円~333円~'[i]
 c=yield max(sorted(p),key=p.get)
 for i in r:e=d[s[:i]];e[c]=e.get(c,1)+1
 s=b'%c'%c+s[:15]

This uses a weighted Bayesian combination of the order 0, ..., 16 Markov models, with weights [1, 6, 12, 30, 65, 99, 87, 117, 118, 89, 95, 118, 96, 184, 126, 219, 126].

The result is not very sensitive to the selection of these weights, but I optimized them because I could, using the same late acceptance hill-climbing algorithm that I used in my answer to "Put together a Senate majority" , where each candidate mutation is just a ±1 increment to a single weight.

Test code

with open('whale2.txt', 'rb') as f:
 g = p()
 wrong = 0
 a = next(g)
 for b in f.read():
 wrong += a != b
 a = g.send(b)
 print(wrong)

Python 3, 2·265 + 510291 = 510821

Predictor

def p():
 d={};s=b''
 while 1:
 p={0:1};r=range(len(s)+1)
 for i in r:
 for c,n in d.setdefault(s[:i],{}).items():p[c]=p.get(c,1)*n**b'1円6円\r#Ihb~jzS177円M210円o{5'[i]
 c=yield max(sorted(p),key=p.get)
 for i in r:e=d[s[:i]];e[c]=e.get(c,1)+1
 s=b'%c'%c+s[:15]

This uses a weighted Bayesian combination of the order 0, ..., 16 Markov models, with weights [1, 6, 13, 35, 73, 104, 98, 126, 106, 122, 83, 127, 77, 136, 111, 123, 53].

Test code

with open('whale.txt', 'rb') as f:
 g = p()
 wrong = 0
 a = next(g)
 for b in f.read():
 wrong += a != b
 a = g.send(b)
 print(wrong)

Python 3, 2·267 + 510193 = 510727

Predictor

def p():
 d={};s=b''
 while 1:
 p={0:1};r=range(len(s)+1)
 for i in r:
 for c,n in d.setdefault(s[:i],{}).items():p[c]=p.get(c,1)*n**b'1円6円\f36円AcWuvY_v`270円~333円~'[i]
 c=yield max(sorted(p),key=p.get)
 for i in r:e=d[s[:i]];e[c]=e.get(c,1)+1
 s=b'%c'%c+s[:15]

This uses a weighted Bayesian combination of the order 0, ..., 16 Markov models, with weights [1, 6, 12, 30, 65, 99, 87, 117, 118, 89, 95, 118, 96, 184, 126, 219, 126].

Test code

with open('whale2.txt', 'rb') as f:
 g = p()
 wrong = 0
 a = next(g)
 for b in f.read():
 wrong += a != b
 a = g.send(b)
 print(wrong)

Python 3, 2·260 + 532465 = 532985

Predictor

def p():
 d={};s=b''
 while 1:
 p={0:1};r=range(len(s)+1)
 for i in r:
 for c,n in d.setdefault(s[:i],{}).items():p[c]=p.get(c,1)*n**b'1円6円16円$I`nnb_YqDrlfG'[i]
 c=yield max(sorted(p),key=p.get)
 for i in r:e=d[s[:i]];e[c]=e.get(c,1)+1
 s=b'%c'%c+s[:15]

This uses a weighted Bayesian combination of the order 0, ..., 16 Markov models, with weights [1, 6, 14, 36, 73, 96, 110, 110, 98, 95, 89, 113, 68, 114, 108, 102, 71].

Test code

with open('whale.txt', 'rb') as f:
 g = p()
 wrong = 0
 a = next(g)
 for b in f.read():
 wrong += a != b
 a = g.send(b)
 print(wrong)

Python 3, 2·265 + 510291 = 510821

Predictor

def p():
 d={};s=b''
 while 1:
 p={0:1};r=range(len(s)+1)
 for i in r:
 for c,n in d.setdefault(s[:i],{}).items():p[c]=p.get(c,1)*n**b'1円6円\r#Ihb~jzS177円M210円o{5'[i]
 c=yield max(sorted(p),key=p.get)
 for i in r:e=d[s[:i]];e[c]=e.get(c,1)+1
 s=b'%c'%c+s[:15]

This uses a weighted Bayesian combination of the order 0, ..., 16 Markov models, with weights [1, 6, 13, 35, 73, 104, 98, 126, 106, 122, 83, 127, 77, 136, 111, 123, 53].

Test code

with open('whale.txt', 'rb') as f:
 g = p()
 wrong = 0
 a = next(g)
 for b in f.read():
 wrong += a != b
 a = g.send(b)
 print(wrong)