Fix uniform mixture model to not depend on undecided trials

Author: mwshinn

pyddm/functions.py

@@ -1011,7 +1011,7 @@ def get_nondecision_time(self, conditions):
         _required_parameters_mixc = list(set(_required_parameters_mixc+[mixture_coef_name]))
     # If mixture coefficient is a parameter or value
     if mixture_distribution is None:
-        mixture_distribution = lambda T : np.ones(len(T))/(np.max(T)+T[1]-T[0])
+        mixture_distribution = lambda T : np.ones(len(T))/(len(T)*(T[1]-T[0]))
     typ, parsed, mixture_dist = _parse_dep(mixture_distribution, "mixture_distribution", ["upper", "T"])
     if typ == "func":
         _required_parameters_mix,_required_conditions_mix,_required_uT_mix = parsed
@@ -1026,7 +1026,6 @@ def apply(self, solution):
                     choice_lower = solution.choice_lower
                     m = solution.model
                     cond = solution.conditions
-                    undec = solution.undec
                     evolution = solution.evolution
                     if _mixc_typ == "val":
                         if isinstance(mixture_coef, Fittable):
@@ -1056,6 +1055,7 @@ def apply(self, solution):
                     assert mix_coef_upper + mix_coef_lower <= 1.001, f"Mixture coefficients cannot be > 1, currently are {mix_coef_upper} and {mix_coef_lower}"
                     choice_upper = choice_upper*(1-(mix_coef_upper+mix_coef_lower)) + lapses_upper*mix_coef_upper*m.dt
                     choice_lower = choice_lower*(1-(mix_coef_upper+mix_coef_lower)) + lapses_lower*mix_coef_lower*m.dt
+                    undec = None if solution.undec is None else solution.undec*(1-(mix_coef_upper+mix_coef_lower))
                     return Solution(choice_upper, choice_lower, m, cond, undec, evolution)
             overlayobjs.append(OverlayMixtureEasy(**{fname:fval for fname,fval in _fittables.items() if fname in list(set(_required_parameters_mix+_required_parameters_mixc))}))
         else:

pyddm/models/overlay.py

@@ -207,14 +207,10 @@ def apply(self, solution):
         choice_lower = solution.choice_lower
         m = solution.model
         cond = solution.conditions
-        undec = solution.undec
+        undec = None if solution.undec is None else solution.undec*(1-self.umixturecoef)
         evolution = solution.evolution
-        # To make this work with undecided probability, we need to
-        # normalize by the sum of the decided density.  That way, this
-        # function will never touch the undecided pieces.
-        norm = np.sum(choice_upper)+np.sum(choice_lower)
-        choice_upper = choice_upper*(1-self.umixturecoef) + .5*self.umixturecoef/len(m.t_domain())*norm
-        choice_lower = choice_lower*(1-self.umixturecoef) + .5*self.umixturecoef/len(m.t_domain())*norm
+        choice_upper = choice_upper*(1-self.umixturecoef) + .5*self.umixturecoef/len(m.t_domain())
+        choice_lower = choice_lower*(1-self.umixturecoef) + .5*self.umixturecoef/len(m.t_domain())
         return Solution(choice_upper, choice_lower, m, cond, undec, evolution)
 
 @paranoidclass

unit_tests.py

@@ -31,10 +31,13 @@ def setUp(self):
         """Create fake models which act like models but are actually much simpler."""
         # Fake model which solves to be a uniform distribution
         class FakeUniformModel(ddm.Model):
+            def __init__(self, *args, p_undecided=.2, **kwargs):
+                self.p_undecided = p_undecided
+                return super().__init__(*args, **kwargs)
             def solve(self, conditions={}, *args, **kwargs):
-                choice_upper = self.t_domain()*0+.4/len(self.t_domain())
-                choice_lower = self.t_domain()*0+.4/len(self.t_domain())
-                undec = self.x_domain(conditions=conditions)*0+.2/len(self.x_domain(conditions=conditions))
+                choice_upper = self.t_domain()*0+(1-self.p_undecided)/2/len(self.t_domain())
+                choice_lower = self.t_domain()*0+(1-self.p_undecided)/2/len(self.t_domain())
+                undec = self.x_domain(conditions=conditions)*0+self.p_undecided/len(self.x_domain(conditions=conditions))
                 return ddm.Solution(choice_upper, choice_lower, self, conditions, undec)
         FakeUniformModel.solve_analytical = FakeUniformModel.solve
         FakeUniformModel.solve_numerical = FakeUniformModel.solve
@@ -213,14 +216,9 @@ def test_OverlayUniformMixture(self):
         s = ddm.Model(drift=ddm.models.DriftConstant(drift=2), noise=ddm.models.NoiseConstant(noise=3)).solve()
         smix = ddm.models.OverlayUniformMixture(umixturecoef=1).apply(s)
         assert np.isclose(np.sum(smix.choice_upper) + np.sum(smix.choice_lower), 1, atol=1e-4)
-        # Should not change uniform distribution
-        s = self.FakeUniformModel(dt=.001).solve()
+        # Should not change uniform distribution if no undecideds
+        s = self.FakeUniformModel(p_undecided=0, dt=.001).solve()
         assert s == ddm.models.OverlayUniformMixture(umixturecoef=.2).apply(s)
-        # Don't change total probability
-        s = ddm.Model(drift=ddm.models.DriftConstant(drift=1)).solve()
-        smix = ddm.models.OverlayUniformMixture(umixturecoef=.2).apply(s)
-        assert np.isclose(np.sum(s.choice_upper) + np.sum(s.choice_lower),
-                          np.sum(smix.choice_upper) + np.sum(smix.choice_lower))
 
     def test_OverlayPoissonMixture(self):
         """Poisson mixture model overlay: an exponential distribution plus the model's solved distribution"""