Traditionally, deep learning is applied to feed-forward tasks, like classification, where the output of the network doesn’t affect the input to the network. It is a decidedly harder problem when the output is recurrently connected such that network output affects the input. Because of this application of deep learning methods to control was largely unexplored until a few years ago. Recently, however, there’s been a lot of progress and research in this area. In this post I’m going to talk about an implementation of deep learning for control presented by Dr. Ilya Sutskever in his thesis Training Recurrent Neural Networks.

In his thesis, Dr. Sutskever uses augmented Hessian-free (AHF) optimization for learning. There are a bunch of papers and posts that go into details about AHF, here’s a good one by Andrew Gibiansky up on his blog, that I recommend you check out. I’m not going to really talk much here about what AHF is specifically, or how it differs from other methods, if you’re unfamiliar there are lots of places you can read up on it. Quickly, though, AHF is kind of a bag of tricks you can use with a fast method for estimating the curvature of the loss function with respect to the weights of a neural network, as well as the gradient, which allows it to make larger updates in directions where the loss function doesn’t change quickly. So rather than estimating the gradient and then doing a small update along each dimension, you can make the size of your update large in directions that change slowly and small along dimensions where things change quickly. And now that’s enough about that.

In this post I’m going to walk through using a Hessian-free optimization library (version 0.3.8) written by my compadre Dr. Daniel Rasmussen to train up a neural network to train up a 2-link arm, and talk about the various hellish gauntlets you need run to get something that works. Whooo! The first thing to do is install this Hessian-free library, linked above.

I’ll be working through code edited a bit for readability, to find the code in full you can check out the files up on my GitHub.

**Build the network**

Dr. Sutskever specified the structure of the network in his thesis to be 4 layers: 1) a linear input layer, 2) 100 Tanh nodes, 3) 100 Tanh nodes, 4) linear output layer. The network is connected up with the standard feedforward connections from 1 to 2 to 3 to 4, plus recurrent connections on 2 and 3 to themselves, plus a ‘skip’ connection from layer 1 to layer 3. Finally, the input to the network is the target state for the plant and the current state of the plant. So, lots of recursion! Here’s a picture:

The output layer connects in to the plant, and, for those unfamiliar with control theory terminology, ‘plant’ just means the system that you’re controlling. In this case an arm simulation.

Before we can go ahead and set up the network that we want to train, we also need to specify the loss function that we’re going to be using during training. The loss function in Ilya’s thesis is a standard one:

where is the total cost of the trajectory generated with , the set of network parameters, is the immediate state cost, is the final state cost, is the state of the arm, is the target state of the arm, is the control signal (torques) that drives the arm, and is a gain value.

To code this up using the `hessianfree`

library we do:

from hessianfree import RNNet from hessianfree.nonlinearities import (Tanh, Linear, Plant) from hessianfree.loss_funcs import SquaredError, SparseL2 l2gain = 10e-3 * dt # gain on control signal loss rnn = RNNet( # specify the number of nodes in each layer shape=[num_states * 2, 96, 96, num_states, num_states], # specify the function of the nodes in each layer layers=[Linear(), Tanh(), Tanh(), Linear(), plant], # specify the layers that have recurrent connections rec_layers=[1,2], # specify the connections between layers conns={0:[1, 2], 1:[2], 2:[3], 3:[4]}, # specify the loss function loss_type=[ # squared error between plant output and targets SquaredError(), # penalize magnitude of control signal (output of layer 3) SparseL2(l2gain, layers=[3])], load_weights=W, use_GPU=True)

Note that if you want to run it on your GPU you’ll need `PyCuda`

and `sklearn`

installed. And a GPU.

An important thing to note as well is that in Dr. Sustkever’s thesis when we’re calculating the squared error of the distance from the arm state to the target, this is measured in joint angles. So it’s kind of a weird set up to be looking at the movement of the hand but have your cost function in joint-space instead of end-effector space, but it definitely simplifies training by making the cost more directly relatable to the control signal. So we need to calculate the joint angles of the arm that will have the hand at different targets around a circle. To do this we’ll take advantage of our inverse kinematics solver from way back when, and use the following code:

def gen_targets(arm, n_targets=8, sig_len=100): #Generate target angles corresponding to target #(x,y) coordinates around a circle import scipy.optimize x_bias = 0 if arm.DOF == 2: y_bias = .35 dist = .075 elif arm.DOF == 3: y_bias = .5 dist = .2 # set up the reaching trajectories around circle targets_x = [dist * np.cos(theta) + x_bias \ for theta in np.linspace(0, np.pi*2, 65)][:-1] targets_y = [dist * np.sin(theta) + y_bias \ for theta in np.linspace(0, np.pi*2, 65)][:-1] joint_targets = [] for ii in range(len(targets_x)): joint_targets.append(arm.inv_kinematics(xy=(targets_x[ii], targets_y[ii]))) targs = np.asarray(joint_targets) # repeat the targets over time for ii in range(targs.shape[1]-1): targets = np.concatenate( (np.outer(targs[:, ii], np.ones(sig_len))[:, :, None], np.outer(targs[:, ii+1], np.ones(sig_len))[:, :, None]), axis=-1) targets = np.concatenate((targets, np.zeros(targets.shape)), axis=-1) # only want to penalize the system for not being at the # target at the final state, set everything before to np.nan targets[:, :-1] = np.nan return targets

And you can see in the last couple lines that to implement the distance to target as a final state cost penalty only we just set all of the targets before the final time step equal to `np.nan`

. If we wanted to penalize distance to target throughout the whole trajectory we would just comment that line out.

**Create the plant**

You’ll notice in the code that defines our RNN I set the last layer of the network to be `plant`

, but that that’s not defined anywhere. Let’s talk. There are a couple of things that we’re going to need to incorporate our plant into this network and be able to use any deep learning method to train it. We need to be able to:

- Simulate the plant forward; i.e. pass in input and get back the resulting plant state at the next timestep.
- Calculate the derivative of the plant state with respect to the input; i.e. how do small changes in the input affect the state.
- Calculate the derivative of the plant state with respect to the previous state; i.e. how do small changes in the plant state affect the state at the next timestep.
- Calculate the derivative of the plant output with respect to its state; i.e. how do small changes in the current position of the state affect the output of the plant.

So 1 is easy, we have the arm simulations that we want already, they’re up on my GitHub. Number 4 is actually trivial too, because the output of our plant is going to be the state itself, so the derivative of the output with respect to the state is just the identity matrix.

For 2 and 3, we’re going to need to calculate some derivatives. If you’ve read the last few posts you’ll note that I’m on a finite differences kick. So let’s get that going! Because no one wants to calculate derivatives!

Important note, the notation in these next couple pieces of code is going to be a bit different from my normal notation because they’re matching with the `hessianfree`

library notation, which is coming from a reinforcement learning literature background instead of a control theory background. So, `s`

is the state of the plant, and `x`

is the input to the plant. I know, I know. All the same, make sure to keep that in mind.

# calculate ds0/dx0 with finite differences d_input_FD = np.zeros((x.shape[0], # number of trials x.shape[1], # number of inputs self.state.shape[1])) # number of states for ii in range(x.shape[1]): # calculate state adding eps to x[ii] self.reset_plant(self.prev_state) inc_x = x.copy() inc_x[:, ii] += self.eps self.activation(inc_x) state_inc = self.state.copy() # calculate state subtracting eps from x[ii] self.reset_plant(self.prev_state) dec_x = x.copy() dec_x[:, ii] -= self.eps self.activation(dec_x) state_dec = self.state.copy() d_input_FD[:, :, ii] = \ (state_inc - state_dec) / (2 * self.eps) d_input_FD = d_input_FD[..., None]

Alrighty. First we create a tensor to store the results. Why is it a tensor? Because we’re going to be doing a bunch of runs at once. So our state dimensions are actually `trials x size_input`

. When we then take the partial derivative, we end up with `trials`

many `size_input x size_state`

matrices. Then we increase each of the parameters of the input slightly one at a time and store the results, decrease them all one at a time and store the results, and compute our approximation of the gradient.

Next we’ll do the same for calculating the derivative of the state with respect to the previous state.

# calculate ds1/ds0 d_state_FD = np.zeros((x.shape[0], # number of trials self.state.shape[1], # number of states self.state.shape[1])) # number of states for ii in range(self.state.shape[1]): # calculate state adding eps to self.state[ii] state = np.copy(self.prev_state) state[:, ii] += self.eps self.reset_plant(state) self.activation(x) state_inc = self.state.copy() # calculate state subtracting eps from self.state[ii] state = np.copy(self.prev_state) state[:, ii] -= self.eps self.reset_plant(state) self.activation(x) state_dec = self.state.copy() d_state_FD[:, :, ii] = \ (state_inc - state_dec) / (2 * self.eps) d_state_FD = d_state_FD[..., None]

Great! We’re getting closer to having everything we need. Another thing we need is a wrapper for running our arm simulation. It’s going to look like this:

def activation(self, x): state = [] # iterate through and simulate the plant forward # for each trial for ii in range(x.shape[0]): self.arm.reset(q=self.state[ii, :self.arm.DOF], dq=self.state[ii, self.arm.DOF:]) self.arm.apply_torque(u[ii]) state.append(np.hstack([self.arm.q, self.arm.dq])) state = np.asarray(state) self.state = self.squashing(state)

This is definitely not the fastest code to run. Much more ideally we would put the state and input into vectors and do a single set of computations for each call to `activation`

rather than having that for loop in there. Unfortunately, though, we’re not assuming that we have access to the dynamics equations / will be able to pass in vector states and inputs.

Squashing

Looking at the above code that seems pretty clear what’s going on, except you might notice that last line calling `self.squashing`

. What’s going on there?

The squashing function looks like this:

def squashing(self, x): index_below = np.where(x < -2*np.pi) x[index_below] = np.tanh(x[index_below]+2*np.pi) - 2*np.pi index_above = np.where(x > 2*np.pi) x[index_above] = np.tanh(x[index_above]-2*np.pi) + 2*np.pi return x

All that’s happening here is that we’re taking our input, and doing nothing to it as long as it doesn’t start to get too positive or too negative. If it does then we just taper it off and prevent it from going off to infinity. So running a 1D vector through this function we get:

This ends up being a pretty important piece of the code here. Basically it prevents wild changes to the weights during learning to result in the system breaking down. So the state of the plant can’t go off to infinity and cause an error to be thrown, stopping our simulation. But because the target state is well within the bounds of where the squashing function does nothing, post-training we’ll still be able to use the resulting network to control a system that doesn’t have this fail safe built in. Think of this function as training wheels that catch you only if you start falling.

With that, we no have pretty much all of the parts necessary to begin training our network!

**Training the network**

We’re going to be training this network on the centre-out reaching task, where you start at a centre point and reach to a bunch of target locations around a circle. I’m just going to be re-implementing the task as it was done in Dr. Sutskever’s thesis, so we’ll have 64 targets around the circle, and train using a 2-link arm. Here’s the code that we’ll use to actually run the training:

for ii in range(last_trial+1, num_batches): # train a bunch of batches using the same input every time # to allow the network a chance to minimize things with # stable input (speeds up training) err = rnn.run_batches(plant, targets=None, max_epochs=batch_size, optimizer=HessianFree(CG_iter=96, init_damping=100)) # save the weights to file, track trial and error # err = rnn.error(inputs) err = rnn.best_error name = 'weights/rnn_weights-trial%04i-err%.3f'%(ii, err) np.savez_compressed(name, rnn.W)

**Training your own network**

A quick aside: if you want to run this code yourself, get a real good computer, have an arm simulation ready, the `hessianfree`

Python library installed, and download and run this `train_hf.py`

file. (Note: I used version 0.3.8 of the `hessianfree`

library, which you can install using `pip install hessianfree==0.3.8`

) This will start training and save the weights into a `weights/`

folder, so make sure that that exists in the same folder as `train_hf.py`

. If you want to view the results of the training at any point run the `plot_error.py`

file, which will load in the most recent version of the weights and plot the error so far. If you want to generate an animated plot like I have below run `gen_animation_plots.py`

and then the last command from my post on generating animated gifs.

Another means of seeing the results of your trained up network is to use the controller I’ve implemented in my controls benchmarking suite, which looks for a set of saved weights in the `controllers/weights`

folder, and will load it in and use it to generate command signals for the arm by running it with

python run.py arm2_python ahf reach --dt=1e-2

where you replace `arm2_python`

with whatever arm model you trained your model on. Note the `--dt=1e-2`

flag, that is important because the model was trained with a .01 timestep and things get a bit weird if you suddenly change the dynamics on the controller.

OK let’s look at some results!

**Results**

OK full discretion, these results are not optimizing the cost function we discussed above. They’re implementing a simpler cost function that *only* looks at the final state, i.e. it doesn’t penalize the magnitude of the control signal. I did this because Dr. Sutskever says in his thesis he was able to optimize with just the final state cost using much smaller networks. I originally looked at neurons with 96 neurons in each layer, and it just took forgoddamnedever to run. So after running for 4 weeks (not joking) and needing to make some more changes I dropped the number of neurons and simplified the task.

The results below are from running a network with 32 neurons in each layer controlling this 2-link arm, and took another 4-5 weeks to train up.

Hey that looks good! Not bad, augmented Hessian-free learning, not bad. It had pretty consistent (if slow) decline in the error rate, with a few crazy bumps from which it quickly recovered. Also take note that each training iteration is actually 32 runs, so it’s not 12,50-ish runs it’s closer to 400,000 training runs that it took to get here.

One biggish thing that was a pain was that it turns out that I only trained the neural network for reaching in the one direction, and when you only train it to reach one way it doesn’t generalize to reaching back to the starting point (which, fair enough). But, I didn’t realize this until I was took the trained network and ran it in the benchmarking code, at which point I was not keen to redo all of the training it took to get the neural network to the level of accuracy it was at under a more complicated training set. The downside of this is that even though I’ve implemented a controller that takes in the trained network and uses it to control the arm, to do the reaching task I have to just do a hard reset after the arm reaches the target, because it can’t reach back to the center, like all the other controllers. All the same, here’s an animation of the trained up AHF controller reaching to 8 targets (it was trained on all 64 above though):

Things don’t always go so smoothly, though. Here’s results from another training run that took around 2-3 weeks, and uses a different 2-link arm model (translated from Matlab code written by Dr. Emo Todorov):

What I found frustrating about this was that if you look at the error over time then this arm is doing as well or better than the previous arm at a lot of points. But the corresponding trajectories look terrible, like something you would see in a horror movie based around getting good machine learning results. This of course comes down to how I specified the cost function, and when I looked at the trajectories plotted over time the velocity of the arm is right at zero at the final time step, which it is not quiiiitte the case for the first controller. So this second network has found a workaround to minimize the cost function I specified in a way I did not intend. To prevent this, doing something like weighting the distance to target heavier than non-zero velocity would probably work. Or possibly just rerunning the training with a different random starting point you could get out a better controller, I don’t have a great feel for how important the random initialization is, but I’m hoping that it’s not all too important and its effects go to zero with enough training. Also, it should be noted I’ve run the first network for 12,500 iterations and the second for less than 6,000, so I’ll keep letting them run and maybe it will come around. The first one looked pretty messy too until about 4,000 iterations in.

**Training regimes**

Frustratingly, the way that you train deep networks is very important. So, very much like the naive deep learning network trainer that I am, I tried the first thing that pretty much anyone would try:

- run the network,
- update the weights,
- repeat.

This is what I’ve done in the results above. And it worked well enough in that case.

If you remember back to the iLQR I made a little while ago, I was able to change the cost function to be

(i.e. to include a penalty for distance to target throughout the trajectory and not just at the final time step) which resulted in straighter trajectories when controlling the 2-link arm. So I thought I would try this here as well. Sadly (incredibly sadly), this was fairly fruitless. The network didn’t really learn or improve much at all.

After much consideration and quandary on my part, I talked with Dr. Dan and he suggested that I try another method:

- run the network,
- record the input,
- hold the input constant for a few batches of weight updating,
- repeat.

This method gave much better results. BUT WHY? I hear you ask! Good question. Let me give giving explanation a go.

Essentially, it’s because the cost function is more complex now. In the first training method, the output from the plant is fed back into the network as input at every time step. When the cost function was simpler this was OK, but now we’re getting very different input to train on at every iteration. So the system is being pulled in different directions back and forth at every iteration. In the second training regime, the same input is given several times in a row, which let’s the system follow the same gradient for a few training iterations before things change again. In my head I picture this as giving the algorithm a couple seconds to catch its breath dunking it back underwater.

This is a method that’s been used in a bunch of places recently. One of the more high-profile instances is in the results published from DeepMind on deep RL for control and for playing Go. And indeed, it also works well here.

To implement this training regime, we set up the following code:

for ii in range(last_trial+1, num_batches): # run the plant forward once rnn.forward(input=plant, params=rnn.W) # get the input and targets from above rollout inputs = plant.get_vecs()[0].astype(np.float32) targets = np.asarray(plant.get_vecs()[1], dtype=np.float32) # train a bunch of batches using the same input every time # to allow the network a chance to minimize things with # stable input (speeds up training) err = rnn.run_batches(inputs, targets, max_epochs=batch_size, optimizer=HessianFree(CG_iter=96, init_damping=100)) # save the weights to file, track trial and error # err = rnn.error(inputs) err = rnn.best_error name = 'weights/rnn_weights-trial%04i-err%.3f'%(ii, err) np.savez_compressed(name, rnn.W)

So you can see that we do one rollout with the weights, then go in and get the `inputs`

and `targets`

that were used in that rollout, and start training the network while holding those constant for `batch_size`

epochs (training sessions). From a little bit of messing around I’ve found `batch_size=32`

to be a pretty good number. So then it runs 32 training iterations where it’s updating the weights, and then saves those weights (because we want a loooottttt of check-points) and then restarts the loop.

Embarrassingly, I’ve lost my simulation results from this trial, somehow…so I don’t have any nice plots to back up the above, unfortunately. But since this is just a blog post I figured I would at least talk about it a little bit, since people might still find it useful if they’re just getting into the field like me. and just update this post whenever I re-run them. If I rerun them.

What I *do* have, however, are results where this method doesn’t work! I tried this with the simpler cost function, that only looks at the final state distance from the target, and it did not go so well. Let’s look at that one!

My guess here is basically that the system has gotten to a point where it’s narrowed things down in the parameter space and now when you run 32 batches it’s overshooting. It needs feedback about its updates after every update at this point. That’s my guess, at least. So it could be the case that for more complex cost functions you’d want to train it while holding the input constant for a while, and then when the error starts to plateau switch to updating the input after every parameter update.

**Conclusions**

All in all, AHF for training neural networks in control is pretty awesome. There are of course still some major hold-backs, mostly related to how long it takes to train up a network, and having to guess at effective training regimes and network structures etc. But! It was able to train up a relatively small neural network to move an arm model from a center point to 64 targets around a circle, with no knowledge of the system under control at all. In Dr. Sutskever’s thesis he goes on to use the same set up under more complicated circumstances, such as when there’s a feedback delay, or a delay on the outgoing control signal, and unexpected noise etc, so it is able to learn under a number of different, fairly complex situations. Which is pretty slick.

Related to the insane training time required, I very easily could be missing some basic thing that would help speed things up. If you, reader, get ambitious and run the code on your own machine and find out useful methods for speeding up the training please let me know! Personally, my plan is to next investigate guided policy search, which seems like it’s found a way around this crazy training time.

Hi Travis!

Great tutorial on using neural networks for control. Thanks for sharing. Looking forward for the Guided Policy Search tutorial.

Thanks! Hoping it won’t be tooooo long before I have one up on GPS! 😀

Nice Post! Looking forward for GPS!

I thought you did this to contrast it with Adaptive Control (the topic of your thesis) and found it lacking, however the conclusion of this post seems rather positive. What’s the difference between Adaptive Control and this approach? Is that a topic for another post?

Good question! The main difference between the adaptive control I discuss in my thesis and the work here is that the adaptive control is able to learn online to compensate for unknown forces, without any background training. In Sutskever’s thesis, he’s able to train up a neural network that is robust to perturbations, but it required training under similar perturbations. Adaptive control _can_ use information about the kind of perturbations that it should expect, but doesn’t require it. And actually I’m hoping to have a post on that in the near future! (As soon as the related paper comes out)

My biggest DL complaint is the insane overhead in training (weeks), so I haven’t actually been able to much direct comparison on the benchmark tasks in my control repo just because it takes so long to train the deep network up to a good level of performance. But the deep network is still able to _eventually_ learn to control a plant it has zero information about to start off with, which is awesome, and is especially cool in situations with plants where controller design isn’t well explored. So one thing I’m pretty interested in exploring (after speeding up training) is combining trained DL networks with an adaptive control component to make them more robust to unexpected perturbations.

Hi Travis

If there is a mismatch in the training plant-model and the testing plant-model will the trained RNN be still robust? or should we make the RNN to learn the change in plant-model. Also Can we use this RNN to train the system online? Thank you. Really an useful post.

Hi Steven,

It very much depends on how close the plants are! In the system that I trained up you can still get decent results with the other two link arm models, but definitely not as clean. If you were to train up the network using both kinds of plants then you should be able to develop something robust to these kinds of changes!

Ideally though you’ll be testing under the same system that you’re training on.

When you say do the training online you mean updating the weights after each run, or as the arm is running? Currently the system updates after every run of the arm, but if you wanted to update the weights as the arm is moving (rather than at the end of the movement) it miiiiiggght be possible but you’ll probably get very poor results!

Glad you enjoyed the post, let me know if you get the chance to mess around with things!

Hi Travis

Thank you for your answers. I set up all the codes required for simulating this. I would like to know if the weights in ‘weights’ folder in controller section is a trained weight or a weight for initialization. I tried using the weights from your controller, unfortunately when I ran run.py the two arm made a zig-zag circular trajectory around the centre. Also, the training period you mentioned in this post (3-4 weeks) is using CPU or GPU? Thanks again for the great post. I am thinking to test this in a continous environment where process states and our target are continuous with time.

Hi Steven,

great! The ‘weights’ file in the weights folder up on my github is post-training. Did you run with the `python run.py arm2_python ahf reach –dt=1e-2` command? Something to be careful about is that it’s been trained to reach the target in .4 seconds, and then if you want to reach to another target you have to reset the system. The easiest way to do this would be set up a script calling the run.py file with all of the above parameters plus –end_time=.4. That’s what I used to generate the animation above. Please let me know if you’re still having trouble getting it running!

The training period I mentioned is when running on my CPU.

Your project sounds interesting! Although could you elaborate more on setting up process states and targets are continuous with time? I’m not sure I follow!

Hi Travis

Thanks for responding all the questions. I was able to run the run.py correctly after specifying the –end_time=0.4 parameter. I tried running train_hf.py with GPU but I could not get tremendous improvement interms of training speed (its little faster).

Continuous I meant, Lets consider a system, say water heater system and we are interested to maintain the temperature of the water at target, say 25 deg C. If the initial water temperature is 20 deg C, here we want to control the water temperature at each instant to take the water temperature from 20 to 25. Even after that we have to continuously control at each and every instant (that is what I meant continuous with time) to make sure that it remains at 25 deg C. So, here we have to monitor both states and target with time.

For such system this is what I tried, I considered a simple first order system. I used Model Predictive Control (MPC) to control that system. Then I collected the data from MPC simulation. Later I learnt the behaviour of MPC with LSTM and Neural Networks (lot of parameters!). Later I used the learnt network to control the system and it did pretty well. The drawback was I was not able to explicitly give the cost function as you have given here, it has just copied MPC. You can check, http://stevenspielberg.me/projects/docs/2_lstmsnn.pdf for more information. Also, I can share the code in a readable format if you want. My next step would be to try training with this approach you have discussed in this post (also looking forward for GPS). Thanks again for the post and code.

Hi Steven,

ah glad that you were able to get running. 🙂 Sad about the GPU! You’ll probably find better results / improvement from using the GPU as the size of the network gets bigger, I believe the size it’s set at is 32 neurons in each of the recurrent layers, you’ll probably start seeing quite a difference in speed if you bump that up.

Ah OK, that description makes sense, sorry I thought you were talking about in comparison to the arm control situation. Sounds like an interesting project to compare all of those systems, I’d definitely be interested to hear the results! And you’re welcome, good luck!

[…] blog: https://studywolf.wordpress.com/2016/04/04/deep-learning-for-control-using-augmented-hessian-free-op… […]