def drawBeforeAfterTeach(self, master, student, learned_student, bpPositions):
if self.draw and random.random() < 1.0/(2.0*float(self.popSize)):
paths = []
endpts = []
maze_lib.SetVerbosity(True)
dataMaster = maze_lib.EvalNetwork(master)
dataBefore = maze_lib.EvalNetwork(student)
dataAfter = maze_lib.EvalNetwork(learned_student)
maze_lib.SetVerbosity(False)
paths.append(dataMaster[3:])
paths.append(dataBefore[3:])
paths.append(dataAfter[3:])
endpts.append((dataMaster[1], dataMaster[2]))
endpts.append((dataBefore[1], dataBefore[2]))
endpts.append((dataAfter[1], dataAfter[2]))
print "drawing before and after backprop learning"
testmaze.drawMaze(self.maze_file, endpts, paths, bpPositions,
self.maze_file + "_" + str(self.numEval) + "_comp",
markerStyle=['.', 'x', '+'], lineColor=['b', 'r', 'g'],
lineLabels=['teacher', 'before', 'after'])
def testGA(nTrials=1, nGen=250, useBackprop=(True, 0.5, 2, 25, False),
maze_file = "medium_maze.txt", draw=True, learningRate=0.05, timesteps=400):
maze_lib.SetUp(learningRate, timesteps)
maze_lib.SetMazePath(os.getcwd() + "/" + maze_file)
maze_lib.SetVerbosity(False)
if not useBackprop or not useBackprop[0]:
useBackprop = False
gen = None
avgBestFit = []
avgMeanFit = []
endpts = []
paths = []
for k in xrange(nTrials):
print "trial number:", k+1
# reload(maze_lib)
# maze_lib.SetUp(learningRate)
# maze_lib.SetMazePath(os.getcwd() + "/" + maze_file)
# maze_lib.SetVerbosity(False)
data = []
ga = genetic(genomeSize=maze_lib.GetNetworkWeightCount(),
hiddenSize=maze_lib.GetNetworkHiddenUnitCount(),
maxEvals=1e308, draw=draw, useBackprop=useBackprop,
maze_file=maze_file)
for i in xrange(nGen):
ga.step()
if ga.numGen % 5 == 0:
data.append(ga.printStats(printGenome=False))
maze_lib.SetVerbosity(True)
print ga.bestGenome
results = maze_lib.EvalNetwork(ga.bestGenome)
maze_lib.SetVerbosity(False)
fitness = results[0]
paths.append(results[3:])
endpts.append((results[1], results[2]))
if not gen:
gen = [x[0] for x in data]
avgBestFit.append([x[2] for x in data])
avgMeanFit.append([x[3] for x in data])
if draw:
return
avgBestFit = np.mean(np.vstack(avgBestFit), axis=0)
avgMeanFit = np.mean(np.vstack(avgMeanFit), axis=0)
plt.figure(figsize=(14,10))
plt.clf()
# plt.ylim(ymin=0)
plt.plot(gen, avgBestFit, label="best fit")
plt.plot(gen, avgMeanFit, label="avg fit")
plt.xlabel("generation")
plt.ylabel("fitness")
plt.legend(loc='lower right')
plt.grid()
out_str = str(time.time()) + "_" + maze_file + "_bp" + \
str(useBackprop) + "_t" + str(timesteps)
plt.savefig(out_str + "_graph.png", bbox_inches="tight", dpi=200)
testmaze.drawMaze(maze_file, endpts, paths, outfile=out_str + "_vis")
testmaze.drawMaze(maze_file, endpts, None, outfile=out_str + "_vis2")
