1 ;;;; XLISP-STAT 2.1 Copyright (c) 1990, by Luke Tierney
2 ;;;; Additions to Xlisp 2.1, Copyright (c) 1989 by David Michael Betz
3 ;;;; You may give out copies of this software; for conditions see the file
4 ;;;; COPYING included with this distribution.
11 (in-package 'lisp-stat
)
13 (export '(bayes-model bayes-model-proto bayes-internals
))
16 ;;;; Objects Representing Functions
19 ;; Generic C2 Functions
21 (defproto c2-function-proto
'(f h num-derivs
))
23 (defmeth c2-function-proto
:isnew
(f &optional
(h .001) (num-derivs 0))
24 (setf (slot-value 'f
) f
)
25 (setf (slot-value 'h
) (if (numberp h
) (list h h
) h
))
26 (setf (slot-value 'num-derivs
) num-derivs
))
28 (defmeth c2-function-proto
:f
(&optional f
)
29 (if f
(setf (slot-value 'f
) f
))
32 (defmeth c2-function-proto
:grad-h
() (first (slot-value 'h
)))
33 (defmeth c2-function-proto
:hess-h
() (second (slot-value 'h
)))
34 (defmeth c2-function-proto
:num-derivs
() (slot-value 'num-derivs
))
36 (defmeth c2-function-proto
:value
(x)
37 (let ((f (send self
:f
)))
40 (let ((v (funcall f x
)))
41 (if (consp v
) (first v
) v
)))))
43 (defmeth c2-function-proto
:gradient
(x &optional
(h (send self
:grad-h
)))
44 (let ((f (send self
:f
)))
45 (if (objectp f
) (send f
:gradient x h
) (numgrad f x nil h
))))
47 (defmeth c2-function-proto
:hessian
(x &optional
(h (send self
:hess-h
)))
48 (let ((f (send self
:f
)))
49 (if (objectp f
) (send f
:hessian x h
) (numhess f x nil h
))))
51 (defmeth c2-function-proto
:vals
(x &optional
(h (send self
:hess-h
)))
52 (let ((f (send self
:f
)))
55 (let ((v (funcall f x
)))
59 (list (first v
) (second v
) (send self
:hessian x h
)))
60 (list v
(send self
:gradient x h
) (send self
:hessian x h
)))))))
62 (defmeth c2-function-proto
:vals
(x &optional
(h (send self
:hess-h
)))
63 (let ((f (send self
:f
)))
64 (if (objectp f
) (send f
:vals x h
) (numhess f x nil h t
))))
67 ;; Scaled C2 Functions
69 (defproto scaled-c2-function-proto
'(scaling) () c2-function-proto
)
71 ;;**** allow function objects?
72 (defmeth scaled-c2-function-proto
:isnew
(f &optional
78 (let* ((value (funcall f theta
))
79 (num-derivs (if (consp value
) (- (length value
) 1) -
1))
80 (sigma-t (if (< 0 num-derivs
) (transpose sigma
))))
84 (-1 (/ (- v center
) scale
))
85 (0 (/ (- (first v
) center
) scale
))
86 (1 (list (/ (- (first v
) center
) scale
)
87 (matmult sigma-t
(/ (second v
) scale
))))
88 (2 (list (/ (- (first v
) center
) scale
)
89 (matmult sigma-t
(/ (second v
) scale
))
90 (matmult sigma-t
(/ (third v
) scale
) sigma
))))))
91 (sf (x) (scale (funcall f
(ax+y sigma x theta t
)))))
92 (call-next-method #'sf h num-derivs
))))
94 ;; Tilted C2 Functions
95 ;; **** allow nil values?
96 (defproto tilt-function-proto
'(tilt exptilt
) () c2-function-proto
)
98 (defmeth tilt-function-proto
:isnew
(&optional f
(tilt .1) (h .001))
99 (call-next-method f h
)
100 (setf (slot-value 'exptilt
) t
)
101 (setf (slot-value 'tilt
) tilt
))
103 (defmeth tilt-function-proto
:tilt
(&optional tilt
)
104 (if tilt
(setf (slot-value 'tilt
) tilt
))
107 (defmeth tilt-function-proto
:exptilt
(&optional
(new nil set
))
108 (if set
(setf (slot-value 'exptilt
) new
))
109 (slot-value 'exptilt
))
111 (defmeth tilt-function-proto
:value
(x)
112 (let ((f (send self
:f
))
113 (tilt (send self
:tilt
))
114 (exptilt (send self
:exptilt
)))
116 (let ((v (send f
:value x
)))
117 (if exptilt v
(log v
)))))
118 (* tilt
(if (consp f
) (apply #'+ (mapcar #'value f
)) (value f
))))))
120 (defmeth tilt-function-proto
:gradient
(x &optional
(h (send self
:grad-h
)))
121 (let ((f (send self
:f
))
122 (tilt (send self
:tilt
))
123 (exptilt (send self
:exptilt
)))
126 (send f
:gradient x h
)
127 (let ((v (send f
:value x
))
128 (grad (send f
:gradient x h
)))
131 (if (consp f
) (apply #'+ (mapcar #'gradient f
)) (gradient f
))))))
133 (defmeth tilt-function-proto
:hessian
(x &optional
(h (send self
:hess-h
)))
134 (let ((f (send self
:f
))
135 (tilt (send self
:tilt
))
136 (exptilt (send self
:exptilt
)))
138 (let* ((vals (send f
:vals x h
))
140 (grad (if exptilt
(second vals
) (/ (second vals
) v
)))
141 (hess (if exptilt
(third vals
) (/ (third vals
) v
))))
142 (if exptilt hess
(- hess
(outer-product grad grad
))))))
143 (* tilt
(if (consp f
) (apply #'+ (mapcar #'hessian f
)) (hessian f
))))))
145 (defmeth tilt-function-proto
:vals
(x &optional
(h (send self
:hess-h
)))
146 (let ((f (send self
:f
))
147 (tilt (send self
:tilt
))
148 (exptilt (send self
:exptilt
)))
150 (let ((vals (send f
:vals x h
)))
153 (let* ((v (first vals
))
154 (grad (/ (second vals
) v
))
155 (hess (- (/ (third vals
) v
)
156 (outer-product grad grad
))))
157 (list (log v
) grad hess
))))))
158 (let ((v (if (consp f
) (mapcar #'vals f
) (vals f
))))
159 (* tilt
(if (consp f
) (apply #'+ v
) v
))))))
161 ;; scaled log posterior prototype
163 (defproto scaled-logpost-proto
164 '(tilt-object init-pars
) () scaled-c2-function-proto
)
166 (defmeth scaled-logpost-proto
:isnew
(f &optional
168 (center 0) (scale 1) (h .001))
169 (let* ((n (length theta
))
171 (m-grad (repeat 0 n
))
172 (m-hess (- (identity-matrix n
)))
173 (pars (list m m-grad m-hess
)))
174 (call-next-method f theta sigma center scale h
)
175 (setf (slot-value 'init-pars
) pars
)
176 (setf (slot-value 'tilt-object
) (send tilt-function-proto
:new
))))
178 (defmeth scaled-logpost-proto
:log-laplace
(g &optional
179 (count-limit 2) det-only
(h .1))
180 (let* ((x (send self
:tilt-newton g count-limit
))
181 (vals (send self
:vals x h
))
182 (gvals (if g
(send g
:vals x h
)))
183 (hess (if g
(+ (third vals
) (third gvals
)) (third vals
)))
184 (det (- (sum (log (diagonal (first (chol-decomp (- hess
)))))))))
187 (if g
(+ (first vals
) (first gvals
) det
) (+ (first vals
) det
)))))
189 (defmeth scaled-logpost-proto
:tilt-newton
(tilt &optional
(count-limit 2))
190 (let* ((pars (slot-value 'init-pars
))
192 (mode-grad (second pars
))
193 (mode-hess (third pars
)))
194 (flet ((gradhess (x initial
)
195 (let ((gh (if (and initial mode-grad mode-hess
)
196 (list mode-grad mode-hess
)
197 (rest (send self
:vals x
)))))
198 (if tilt
(+ gh
(rest (send tilt
:vals x
))) gh
)))
199 (newton-step (x gh
) (- x
(solve (second gh
) (first gh
)))))
200 (do* ((count 1 (+ count
1))
201 (gradhess (gradhess mode t
) (gradhess x nil
))
202 (x (newton-step mode gradhess
) (newton-step x gradhess
)))
203 ((>= count count-limit
) x
)))))
205 (defmeth scaled-logpost-proto
:tilt-laplace
(g tilt
&optional
206 (exptilt t
) maxiter det-only h
)
207 (let ((tilt-object (slot-value 'tilt-object
)))
208 (send tilt-object
:exptilt exptilt
)
209 (send tilt-object
:f g
)
210 (send tilt-object
:tilt tilt
)
211 (send self
:log-laplace tilt-object maxiter det-only h
)))
213 (defmeth scaled-logpost-proto
:tilt-mode
(g tilt
&key
(exptilt t
) (maxiter 2))
214 (let ((tilt-object (slot-value 'tilt-object
)))
215 (send tilt-object
:exptilt exptilt
)
216 (send tilt-object
:f g
)
217 (send tilt-object
:tilt tilt
)
218 (send self
:tilt-newton tilt-object maxiter
)))
221 ;;;; Bayes Model Prototype
224 (defproto bayes-model-proto
'(bayes-internals))
226 ;; initialization methods and constructor function
228 (defmeth bayes-model-proto
:isnew
(logpost mode
&key
234 (send self
:set-bayes-internals
235 logpost mode scale derivstep nil nil t domain
)
236 (if maximize
(send self
:maximize verbose
)))
238 (defun bayes-model (logpost mode
&rest args
&key
(quick t
) (print t
))
239 "Args: (logpost mode &key scale derivstep (verbose t)
240 (quick t) (print t)))
241 LOGPOST computes the logposterior density. It should return the
242 function, or a list of the function value and gradient, or a list of
243 the function value, gradient and Hessian. MODE is an initial guess for
244 the mode. SCALE and DERIVSTEP are used for numerical derivatives and
245 scaling. VERBOSE controls printing of iteration information during
246 optimization, PRINT controls printing of summary information. If QUICK
247 is T the summary is based on first order approximations."
248 (let ((m (apply #'send bayes-model-proto
:new logpost mode args
)))
249 (if print
(send m
:display
:quick quick
))
254 (defmeth bayes-model-proto
:display
(&key
(quick t
))
255 (let* ((moments (send self
(if quick
:1stmoments
:moments
)))
256 (means (first moments
))
257 (stdevs (second moments
))
258 (p-names (send self
:parameter-names
)))
260 (format t
"~2%First Order Approximations to Posterior Moments:~2%")
261 (format t
"~2%Approximate Posterior Moments:~2%"))
262 (mapcar #'(lambda (name mu sd
)
263 (format t
"~22a ~10g (~a)~%" name mu sd
))
269 (defmeth bayes-model-proto
:parameter-names
()
270 (let ((n (length (send self
:mode
))))
271 (mapcar #'(lambda (x) (format nil
"Parameter ~d" x
)) (iseq 0 (- n
1)))))
273 ;; implementation-dependent access methods
275 (defmeth bayes-model-proto
:set-bayes-internals
(lp m s h mval ch max dom
)
276 (setf (slot-value 'bayes-internals
)
277 (vector lp m s h mval ch max dom
)))
279 (defmeth bayes-model-proto
:logpost
(&optional new
)
280 (let ((internals (slot-value 'bayes-internals
)))
282 (setf (select internals
0) new
)
283 (send self
:needs-maximizing t
))
284 (select internals
0)))
286 (defmeth bayes-model-proto
:domain
(&optional new
)
287 (let ((internals (slot-value 'bayes-internals
)))
288 (if new
(setf (select internals
7) new
))
289 (select internals
7)))
291 (defmeth bayes-model-proto
:mode-values
(&optional mode mval ch
)
292 (let ((internals (slot-value 'bayes-internals
)))
294 (setf (select internals
1) mode
)
295 (setf (select internals
4) mval
)
296 (setf (select internals
5) ch
))
297 (list (select internals
1)
299 (select internals
5))))
301 (defmeth bayes-model-proto
:parameter-scale
(&optional new
)
302 (let ((internals (slot-value 'bayes-internals
)))
303 (if new
(setf (select internals
2) new
))
304 (select internals
2)))
306 (defmeth bayes-model-proto
:parameter-dimension
()
307 (length (select (slot-value 'bayes-internals
) 1)))
309 (defmeth bayes-model-proto
:derivstep
()
310 (select (slot-value 'bayes-internals
) 3))
312 (defmeth bayes-model-proto
:needs-maximizing
(&optional
(new nil set
))
313 (let ((internals (slot-value 'bayes-internals
)))
314 (if set
(setf (select internals
6) new
))
315 (select internals
6)))
317 ;; Transformation-Related Methods
318 ;; (These should be the only ones needing to be changed to handle
319 ;; an internal parameter transformation; perhaps also :logpost)
321 ;; **** fix to be more careful about use of functionp
322 (defun function-list (g &optional n
)
324 ((or (functionp g
) (objectp g
)) (list g
))
327 (list #'(lambda (x) (elt x g
)))
328 (let ((grad (make-array n
:initial-element
0))
329 (hess (make-array (list n n
) :initial-element
0)))
330 (setf (aref grad g
) 1)
331 (list #'(lambda (x) (list (elt x g
) grad hess
))))))
332 (t (mapcar #'(lambda (x) (car (function-list x n
))) g
))))
334 (defmeth bayes-model-proto
:mode
()
335 (if (send self
:needs-maximizing
) (send self
:maximize
))
336 (first (send self
:mode-values
)))
338 (defmeth bayes-model-proto
:new-mode-guess
(new)
339 (send self
:needs-maximizing t
)
340 (send self
:mode-values new
))
342 (defmeth bayes-model-proto
:transformed-logpost
()
343 (if (send self
:needs-maximizing
) (send self
:maximize
))
344 (let* ((m-values (send self
:mode-values
))
345 (mode (first m-values
))
346 (mval (second m-values
))
347 (ch (third m-values
))
348 (h (send self
:derivstep
))
349 (f (send self
:logpost
)))
350 (send scaled-logpost-proto
:new f mode ch mval
1 h
)))
352 ;;**** need transformed domain here
354 (defmeth bayes-model-proto
:transformed-functions
(&optional g
(c 0) (s 1))
355 (if (send self
:needs-maximizing
) (send self
:maximize
))
356 (let* ((m-values (send self
:mode-values
))
357 (mode (first m-values
))
358 (mval (second m-values
))
359 (ch (third m-values
))
360 (h (send self
:derivstep
))
362 (g (function-list (if g g
(iseq n
)) n
))
363 (c (if (numberp c
) (repeat c
(length g
)) c
))
364 (s (if (numberp s
) (repeat s
(length g
)) s
)))
365 (mapcar #'(lambda (g c s
)
366 (send scaled-c2-function-proto
:new g mode ch c s h
))
371 (defmeth bayes-model-proto
:maximize
(&optional
(verbose 0))
372 (let* ((lp (send self
:logpost
))
373 (x (first (send self
:mode-values
)))
374 (scale (send self
:parameter-scale
))
375 (h (send self
:derivstep
))
376 (minfo (newtonmax lp x
382 (mval (second minfo
))
383 (ch (first (chol-decomp (inverse (- (fourth minfo
)))))))
384 (send self
:mode-values mode mval ch
)
385 (send self
:needs-maximizing nil
)
386 (send self
:check-derivatives verbose
)))
388 (defmeth bayes-model-proto
:check-derivatives
(&optional
391 (let* ((verbose (if (numberp verbose
) (< 0 verbose
) verbose
))
392 (n (send self
:parameter-dimension
))
393 (tlp (send self
:transformed-logpost
))
394 (hess (send tlp
:hessian
(repeat 0 n
)))
395 (needs-max (send self
:needs-maximizing
)))
396 (when (> (max (abs (+ hess
(identity-matrix n
)))) epsilon
)
397 (if verbose
(format t
"Adjusting derivatives...~%"))
398 (let* ((ch (first (chol-decomp (- (inverse hess
)))))
399 (mvals (send self
:mode-values
))
400 (m (matmult (third mvals
) ch
)))
401 (send self
:mode-values
(first mvals
) (second mvals
) m
)
402 (if (not needs-max
) (send self
:needs-maximizing nil
))
404 (let* ((tlp (send self
:transformed-logpost
))
405 (hess (send tlp
:hessian
(repeat 0 n
))))
406 (if (> (max (abs (+ hess
(identity-matrix n
)))) epsilon
)
408 "Derivatives may not be well-behaved.~%"))))))))
412 (defmeth bayes-model-proto
:1stmoments
(&optional gfuns
&key covar
)
413 "Args: (&optional gfuns &key covar)
414 Computes first order approximations to posterior moments. GFUNS can be
415 a parameter index, list of indices, a function of the parameters or a
416 list of such functions. Returns a the list of first order approximate
417 means and standard deviations if COVAR is NIL. If COVAR is T the
418 covaraince is appended to the end of the result as well."
419 (if (send self
:needs-maximizing
) (send self
:maximize
))
420 (let* ((n (send self
:parameter-dimension
))
422 (g (send self
:transformed-functions gfuns
0 1))
423 (grads (apply #'bind-columns
424 (mapcar #'(lambda (g) (send g
:gradient x
)) g
)))
425 (mean (mapcar #'(lambda (g) (send g
:value x
)) g
))
426 (cov (matmult (transpose grads
) grads
)))
428 (list mean
(sqrt (diagonal cov
)) cov
)
429 (list mean
(sqrt (diagonal cov
))))))
431 (defmeth bayes-model-proto
:mgfmoments
(&optional g
&key covar
435 (let* ((moms1 (send self
:1stmoments g
:covar covar
))
436 (mean1 (first moms1
))
437 (stdev1 (second moms1
))
438 (cov1 (if covar
(third moms1
)))
439 (l-object (send self
:transformed-logpost
))
440 (g-objects (send self
:transformed-functions g mean1 stdev1
))
441 (ldet0 (send l-object
:log-laplace nil maxiter t h
)))
442 (labels ((lapdet (g tilt
)
443 (- (send l-object
:tilt-laplace g tilt t maxiter t h
) ldet0
))
445 (let ((ldet1 (lapdet g mgfdel
))
446 (ldet2 (lapdet g
(- mgfdel
))))
447 (list (+ m
(* s
(/ (- ldet1 ldet2
) (* 2 mgfdel
))))
448 (* s
(sqrt (+ 1 (/ (+ ldet1 ldet2
) (^ mgfdel
2))))))))
450 (let* ((mu (first mean-sd
))
451 (sd (second mean-sd
))
452 (cov (diagonal (^ sd
2)))
455 (rvdiff (/ (- var var1
) var
))
457 (2tilt2 (* 2 (^ tilt
2)))
459 (dotimes (i (length g
) cov
)
461 (let* ((g (select g
(list i j
)))
462 (rvdi (select rvdiff i
))
463 (rvdj (select rvdiff j
))
466 (ldt1 (lapdet g tilt
))
467 (ldt2 (lapdet g negtilt
))
468 (del2 (/ (+ ldt1 ldt2
) 2tilt2
))
469 (d (- del2
(* 0.5 rvdi
) (* 0.5 rvdj
)))
470 (c (+ (aref cov1 i j
) (* d sdi sdj
))))
471 (setf (aref cov i j
) c
)
472 (setf (aref cov j i
) c
)))))))
473 (let ((mean-sd (transpose (mapcar #'moms2 mean1 stdev1 g-objects
))))
475 (append mean-sd
(list (covar g-objects mean-sd
)))
478 (defmeth bayes-model-proto
:fullmoments
(&optional g
&key covar
481 (let* ((moms1 (send self
:1stmoments g
))
482 (mean1 (first moms1
))
483 (stdev1 (second moms1
))
484 (l-object (send self
:transformed-logpost
))
485 (g-objects (send self
:transformed-functions g
0 mean1
))
486 (loglap0 (send l-object
:log-laplace nil maxiter nil h
)))
487 (labels ((loglap (g tilt
)
488 (- (send l-object
:tilt-laplace g tilt nil maxiter nil h
)
491 (let ((mu1 (exp (loglap g
1.0)))
492 (mu2 (exp (loglap g
2.0))))
493 (* mu
(list mu1
(sqrt (- mu2
(^ mu1
2)))))))
495 (let* ((mu (/ (first mean-sd
) mean1
))
496 (sd (second mean-sd
))
497 (cov (diagonal (^ sd
2))))
498 (dotimes (i (length g
) cov
)
500 (let* ((g (select g
(list i j
)))
501 (muij (exp (loglap g
1.0)))
504 (mu1i (select mean1 i
))
505 (mu1j (select mean1 j
))
506 (c (* (- muij
(* mui muj
)) mu1i mu1j
)))
507 (setf (aref cov i j
) c
)
508 (setf (aref cov j i
) c
)))))))
509 (let ((mean-sd (transpose (mapcar #'moms2 g-objects mean1
))))
511 (append mean-sd
(list (covar g-objects mean-sd
)))
514 (defmeth bayes-model-proto
:2ndmoments
(&rest args
)
515 (apply #'send self
:mgfmoments args
))
517 (defmeth bayes-model-proto
:moments
(&rest args
)
518 "Args: (&optional gfuns &key covar)
519 Computes second order approximations to posterior moments. GFUNS can be
520 a parameter index, list of indices, a function of the parameters or a
521 list of such functions. Returns a the list of second order approximate
522 means and standard deviations if COVAR is NIL. If COVAR is T the
523 covaraince is appended to the end of the result as well."
524 (apply #'send self
:2ndmoments args
))
528 (defproto laplace-margin-proto
'(logpost g x val i j a grad gval lu h
))
530 (defmeth laplace-margin-proto
:isnew
(logpost g n k h
)
531 (setf (slot-value 'logpost
) logpost
)
532 (setf (slot-value 'g
) g
)
533 (setf (slot-value 'x
) (repeat 0 (+ n k
)))
534 (setf (slot-value 'i
) (iseq n
))
535 (setf (slot-value 'j
) (+ n
(iseq k
)))
536 (setf (slot-value 'a
)
537 (make-array (list (+ n k
) (+ n k
)) :initial-element
0))
538 (setf (slot-value 'h
) h
)
539 (send self
:adjust-internals t
))
541 (defmeth laplace-margin-proto
:adjust-internals
(&optional initial
)
542 (let* ((logpost (slot-value 'logpost
))
550 (lambda (select x j
))
553 (list 0 (repeat 0 n
) (- (identity-matrix n
)))
554 (send logpost
:vals y h
)))
558 (gvals (mapcar #'(lambda (x) (send x
:vals y h
)) g
))
559 (gval (mapcar #'first gvals
))
560 (ggrad (mapcar #'second gvals
))
561 (ghess (mapcar #'third gvals
))
562 (ggradmat (apply #' bind-columns ggrad
)))
563 (setf (slot-value 'val
) val
)
564 (setf (slot-value 'grad
) (apply #'+ grad
(* lambda ggrad
)))
565 (setf (slot-value 'gval
) gval
)
566 (setf (select a i i
) (apply #'+ hess
(* lambda ghess
)))
567 (setf (select a i j
) ggradmat
)
568 (setf (select a j i
) (transpose ggradmat
))
569 (setf (slot-value 'lu
) (lu-decomp a
))))
571 ;; **** test for nonsingularity?
572 (defmeth laplace-margin-proto
:move-to
(target)
573 (let* ((x (slot-value 'x
))
574 (grad (slot-value 'grad
))
575 (gval (slot-value 'gval
))
576 (lu (slot-value 'lu
))
577 (next-x (- x
(lu-solve lu
(combine grad
(- gval target
))))))
578 (setf (slot-value 'x
) next-x
)
579 (send self
:adjust-internals
)))
581 (defmeth laplace-margin-proto
:log-density
(&optional profile
)
582 (let ((val (slot-value 'val
)))
585 (let* ((lu (slot-value 'lu
))
586 (nonsing (null (fourth lu
))))
588 (+ (* -
0.5 (sum (log (abs (diagonal (first lu
))))))
591 ;; ***** fix step choice
592 ;; ***** Cut off at first nil?
593 (defmeth bayes-model-proto
:log-margin1
(g x
&key
597 (let* ((moms1 (send self
:1stmoments g
))
598 (mean1 (select (first moms1
) 0))
599 (stdev1 (select (second moms1
) 0))
600 (n (send self
:parameter-dimension
))
601 (l-ob (send self
:transformed-logpost
))
602 (g-obs (send self
:transformed-functions g mean1 stdev1
))
603 (xs (/ (- x mean1
) stdev1
))
604 (xlow (coerce (sort-data (select xs
(which (<= xs
0)))) 'list
))
605 (xhigh (coerce (sort-data (select xs
(which (> xs
0)))) 'list
)))
607 (let ((margin (send laplace-margin-proto
:new l-ob g-obs n
1 h
)))
608 (flet ((nextmargin (x)
609 (send margin
:move-to x
)
610 (send margin
:log-density profile
)))
611 (mapcar #'nextmargin x
)))))
612 (let* ((ylow (reverse (margin (reverse xlow
))))
613 (yhigh (margin xhigh
))
614 (x (append xlow xhigh
))
615 (y (append ylow yhigh
))
616 (i (which (mapcar #'numberp y
)))
619 (xy (if spline
(spline xi yi
) (list xi yi
))))
620 (list (+ mean1
(* stdev1
(first xy
)))
621 (- (second xy
) (log stdev1
) (* 0.5 (log (* 2 pi
)))))))))
623 (defmeth bayes-model-proto
:margin1
(g x
&key
627 "Args: (g x &key (:derivstep .05) (spline t) profile)
628 Computes Laplace approximation to marginal posterior density of G at
629 points X. G can be an index or a function of the parameter vector. X
630 is a sequence that should include the modal value of G. If SPLINE is
631 true the log density is splined. If PROFILE is true, a profile of the
632 posterior is returned."
633 (let* ((logmar (send self
:log-margin1 g x
637 (list (first logmar
) (exp (second logmar
)))))
639 ;;**** allow domain test function
640 (defmeth bayes-model-proto
:impsample
(&optional g
&key
(n 100) (df 2))
641 (let* ((l-ob (send self
:transformed-logpost
))
642 (g-obs (send self
:transformed-functions g
))
643 (k (send self
:parameter-dimension
))
644 (v (chisq-rand n df
))
645 (z (* (normal-rand (repeat k n
)) (sqrt (/ df v
))))
646 (c (- (log-gamma (/ (+ k df
) 2))
648 (* (/ k
2) (log (/ df
2))))))
650 (let ((lp (send l-ob
:value z
))
651 (lt (* -
0.5 (+ k df
) (log (+ 1 (/ (sum (* z z
)) df
))))))
652 (if (realp lp
) (exp (- lp lt c
)) 0)))
653 (gvals (z) (mapcar #'(lambda (g) (send g
:value z
)) g-obs
)))
654 (list (mapcar #'gvals z
) (mapcar #'w z
)))))
656 (defmeth bayes-model-proto
:impmoments
(&optional g
&key
(n 100) (df 2))
657 (let* ((impsample (send self
:impsample g
:n n
:df df
))
658 (means (/ (reduce #'+ (* (first impsample
) (second impsample
)))
659 (reduce #'+ (second impsample
))))
660 (x (mapcar #'(lambda (z) (^
(- z means
) 2)) (first impsample
)))
661 (vars (/ (reduce #'+ (* x
(second impsample
)))
662 (reduce #'+ (second impsample
)))))
663 (list means
(sqrt vars
))))