DiffOpt
DiffOpt implements differentiable optimal controls in TensorFlow 2. Have Fun! -Brandon
Setup
Install this package using pip.
pip install git+http://github.com/brandontrabucco/diffopt.git
Usage
Collect a batch of initial states to use for planning.
# a tf.EagerTensor with shape [batch_dim, state_dim]
initial_states = tf.random.normal([1, 3])Define an initial policy where the optimization process starts.
# a diffopt.Distribution that returns [batch_dim, controls_dim]
controls_model = diffopt.Zeros(1)Create a dynamics model that predicts future states given current states and controls.
# a diffopt.Distribution that returns [batch_dim, state_dim]
A = tf.constant([[[-0.313, 56.7, 0.0], [-0.0139, -0.426, 0.0], [0.0, 56.7, 0.0]]])
B = tf.constant([[[0.232], [0.0203], [0.0]]])
dynamics_model = diffopt.Linear(0, [0, 0], [A, B])Create a cost model that evaluates the cost of states and controls.
# a diffopt.Distribution that returns [batch_dim, 1]
Q = tf.constant([[[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 1.0]]]])
R = tf.constant([[[[1.0]]]])
cost_model = diffopt.Quadratic(0, [0, 0], [0, 0], [[Q, 0], [0, R]])Launch the optimizer to get a new policy.
# a diffopt.TimeVaryingLinearGaussian that returns [batch_dim, controls_dim]
controls_model = diffopt.iterative_lqr(
initial_states,
controls_model,
dynamics_model,
cost_model,
h=20,
n=10,
a=0.1,
deterministic=True)