# Distinguishing Poisionous From Edible Wild Mushrooms

Document Sample

```					Ray LaBarge
ECE 539 Project

DISTINGUISHING POISONOUS FROM
EDIBLE WILD MUSHROOMS
PROBLEM STATEMENT

   Wild Mushrooms are
available throughout
North America
   No “Leaves of 3, Leave it
Be” rule to classify by
   Consequences of
misclassification are
PRIOR RESEARCH

 Dataset from UC-Irvine Machine Learning
Database
 Most notable work by Prof. Włodzisław Duch of
Poland’s Nicolaus Copernicus University
 Established                  3 classifying rules with 100% accuracy
   R   1   : odor ∈ {a,l,n}→ class e
   R   2   : odor ∈ {a,l,n} & spore-print-color ∈ {r,w} & population = v & habitat ∈ {l,p}→class e
   R   3   : odor ∈ {a,l,n} & spore-print-color ∈ {r,w} & population = v & gill-size = b →class e
   R   4   : ELSE→ class p
APPROACH

   Develop k-NN Algorithm
 Determine   baseline performance
 Establish if any relation between physical
characteristics and toxicity exists
   Develop a Perceptron
 Randomize  data and partition into Train and Test
 Use 3 way cross validation

   All algorithms implemented in Java
99.90% accurate averaged from
RESULTS - KNN   a 3-way cross validation
RESULTS - PERCEPTRON
Training Set Accuracy by Epoch Size and Learning Rate
99.8

99.795

99.79
Training Set Accuracy (%)

99.785

Learning
99.78                                                             Rate
0.001
99.775
0.01
0.1
99.77

99.765
More than 99% accuracy!
99.76

99.755
1                      10                    100
Epoch Size

Percentages are averaged over 3 cross validations of data
RESULTS – PERCEPTRON

Testing Set Accuracy by Epoch Size and Learning Rate
99.77

99.76

99.75
Testing Set Accuracy (%)

99.74
Learning Rate
0.01, 0.001
99.73
0.1

99.72

99.71
More than 99% accuracy!
99.7
1                      10                  100
Epoch Size

Note: Using learning rates of 0.01 and 0.001 had the same testing set accuracy.
MOST LIKELY POISONOUS IF...          MOST LIKELY EDIBLE IF…

1.   Spore print color is            1.   Ring number is
green (0.54 weight)                  two (1.84 weight!!!)
2.   Stalk surface below the ring    2.   Odor is either
is scaly (0.25 weight)               almond (0.43 weight)
3.   Stalk color below the ring is        anise (0.42 weight)
no smell at all (0.42 weight)
3.   Ring type is

RESULTS

```
DOCUMENT INFO
Shared By:
Categories:
Stats:
 views: 43 posted: 8/19/2010 language: English pages: 8