def setUp(self):
self.R = np.asarray([[ 1.0, 0.4, -0.4],
[ 0.4, 1.0, 0.6],
[-0.4, 0.6, 1.0]])
self.mu = [100.0] * 3
self.sigma = [10.0] * 3
self.cov = rewards.cor2cov(self.R, self.sigma)
self.maze = grid.make_multimaze(4, 4, 3)
self.goals = grid.maze_goal_states(self.maze, 3, self.mu, self.cov)
if not args.stdev:
args.stdev = np.ones(args.tasks)
else:
args.stdev = np.asarray(ast.literal_eval(args.stdev))
# compute a covariance matrix from the correlation matrix and standard deviations
cov = rwd.cor2cov(args.correlation, args.stdev)
if not rwd.is_pos_def(cov):
print('Error: covariance matrix must be positive definite', file=sys.stderr)
sys.exit(1)
# maze type instances
if args.type == 'maze':
# TODO: incorporate correlated rewards somehow
maze = grd.make_multimaze(args.rows, args.cols, args.tasks)
goals = grd.maze_goal_states(maze, args.tasks, args.rmeans, cov)
io.write_maze_instance(maze, goals)
print('# type={}, rows={}, cols={}, correlation={}, stdev={}'.
format(args.type, args.rows, args.cols, args.correlation.tolist(), args.stdev.tolist()))
# perturbation of existing instances
elif args.type == 'perturbation':
if not args.baseline:
parser.print_help()
sys.exit(1)
else:
# TODO: handle other kinds of models
(net, trainset) = io.read_neural_net(args.baseline)
net2 = net.fuzz_neural_net(net, args.fuzz_frac, args.fuzz_scale)
