Robot Learning HW2: Model-Based RL Solution

Description

• Introduction

The goal of this assignment is to get experience with model-based reinforcement learning. In general, model-based reinforcement learning consists of two main parts: learning a dynamics function to model observed state transitions, and then using predictions from that model in some way to decide what to do (e.g., use model predictions to learn a policy, or use model predictions directly in an optimization setup to maximize predicted rewards).

In this assignment, you will do the latter. You will implement both the process of learning a dynamics model, as well as the process of creating a controller to perform action selection through the use of these model predictions. For references to this type of approach, see this paper and this paper.

• Model-Based Reinforcement Learning

We will now provide a brief overview of model-based reinforcement learning (MBRL), and the specific type of MBRL you will be implementing in this home-work. Please see the lecture slides for additional details.

MBRL consists primarily of two aspects: (1) learning a dynamics model and (2) using the learned dynamics models to plan and execute actions that minimize a cost function (or maximize a reward function).

2.1 Dynamics Model

which predicts the change in state given the current state and action. So given

ˆ

the prediction ∆_t+1, you can generate the next prediction with

ˆ	(2)
ˆst+1 = st + ∆t+1.

See the previously referenced paper for intuition on why we might want our net-work to predict state differences, instead of directly predicting next state.

You will train f_θ in a standard supervised learning setup, by performing gradient descent on the following objective:

L(θ) =	∥(st+1 − st) − fθ(st, at)∥22		(3)
	(st,at,st+1)∈D
=	ˆ	2	(4)
	∥∆t+1 − ∆t+1	∥2

(s_t,a_t,s_t+1)∈D

In practice, it’s helpful to normalize the target of a neural network. So in the code, we’ll train the network to predict a normalized version of the change in state, as in

L(θ) =	∥Normalize(st+1 − st) − fθ(st, at)∥22.	(5)
	(st,at,st+1)∈D

Since f_θ is trained to predict the normalized state difference, you generate the next prediction with

ˆst+1 = st + Unnormalize(fθ(st, at)).

(6)

2.2 Action Selection

Given the learned dynamics model, we now want to select and execute actions that minimize a known cost function (or maximize a known reward function). Ideally, you would calculate these actions by solving the following optimiza-tion:

at∗ = arg min	∞
	c(ˆst′, at′) where ˆst′+1 = ˆst′ + fθ(ˆst′, at′).	(7)
at:∞	t′=t

However, solving Eqn. 7 is impractical for two reasons: (1) planning over an infinite sequence of actions is impossible and (2) the learned dynamics model is imperfect, so using it to plan in such an open-loop manner will lead to ac-cumulating errors over time and planning far into the future will become very inaccurate.

Instead, one alternative is to solve the following gradient-free optimization prob-lem:

A∗ = arg		t+H−1
	min	c(ˆst′, at′) s.t. ˆst′+1 = ˆst′ + fθ(ˆst′, at′), (8)
	{A(0),…,A(K−1)}	t′=t

in which A^(k) = (a⁽_t^k), . . . , a⁽_t+^k)_H−1) are each a random action sequence of length H. What Eqn. 8 says is to consider K random action sequences of length H, predict the result (i.e., future states) of taking each of these action sequences using the learned dynamics model f_θ, evaluate the cost/reward associated with each candidate action sequence, and select the best action sequence. Note that this approach only plans H steps into the future, which is desirable because it prevent accumulating model error, but is also limited because it may not be sufficient for solving long-horizon tasks.

A better alternative to this random-shooting optimization approach is the cross-entropy method (CEM), which is similar to random-shooting, but with iterative improvement of the distribution of actions that are sampled from. We first randomly initialize a set of K action sequences A⁽⁰⁾, …, A^(K−1), like in random-shooting. Then, we choose the M sequences with the highest predicted sum of discounted rewards as the ”elite” action sequences. We then fit a diagonal Gaussian with the same mean and variance as the ”elite” action sequences, and use this as our action sampling distribution for the next iteration. After repeating this process L times, we take the final mean of the Gaussian as the optimized action sequence. See Section 3.3 in this paper for more details.

Additionally, since our model is imperfect and things will never go perfectly according to plan, we adopt a model predictive control (MPC) approach, where at every time step we perform random-shooting or CEM to select the best H-step action sequence, but then we execute only the first action from that sequence before replanning again at the next time step using updated state information. This reduces the effect of compounding errors when using our approximate dynamics model to plan too far into the future.

2.3 On-Policy Data Collection

Although MBRL is in theory off-policy—meaning it can learn from any data—in practice it will perform poorly if you don’t have on-policy data. In other words, if a model is trained on only randomly-collected data, it will (in most cases) be insufficient to describe the parts of the state space that we may actually care about. We can therefore use on-policy data collection in an iterative algorithm to improve overall task performance. This is summarized as follows:

Algorithm 1 Model-Based RL with On-Policy Data

Run base policy π₀(a_t, s_t) (e.g., random policy) to collect D = {(s_t, a_t, s_t+1)} while not done do

Train f_θ using D (Eqn. 4)

s_t ← current agent state

for rollout number m = 0 to M do

for timestep t = 0 to T do

A^∗ = π_MPC(a_t, s_t) where π_MPC is obtained from random-shooting or CEM

a_t ← first action in A^∗

Execute a_t and proceed to next state s_t+1

Add (s_t, a_t, s_t+1) to D

end

end

end

2.4 Ensembles

A simple and effective way to improve predictions is to use an ensemble of models. The idea is simple: rather than training one network f_θ to make pre-dictions, we’ll train N independently initialized networks {f_θn }^N_n=1, and average their predictions to get your final predictions

	1	N
f(s , a ) =		f	(s	, a ).	(9)
t t		θn	t	t
	N n=1

In this assignment, you’ll train an ensemble of networks and compare how dif-ferent values of N effect the model’s performance.

• Code

You will implement the MBRL algorithm described in the previous section.

3.1 Overview

Obtain the code from https://github.com/milarobotlearningcourse/ ift6163_homeworks/tree/master/hw2.

You will add code to the following three files:agents/mb_agent.py, models/ff_model.py, and policies/MPC_policy.py. You will also need to edit these files by

copying code from past homeworks or Piazza: infrastructure/rl_trainer.py and infrastructure/utils.py.

Problem 1

What you will implement:

Collect a large dataset by executing random actions. Train a neural network dynamics model on this fixed dataset and visualize the resulting predictions. The implementation that you will do here will be for training the dynamics model, and comparing its predictions against ground truth. You will be reusing the utilities you wrote for HW1 (or Piazza) for the data collection part (look for “get this from Piazza” markers).

What code files to fill in:

1. ift6131/agents/mb_agent.py

2. ift6131/models/ff_model.py

3. ift6131/infrastructure/utils.py

4. ift6131/policies/MPC_policy.py (just one line labeled TODO(Q1) for now)

What commands to run:

python run_hw2_mb.py exp_name=q1_cheetah_n500_arch1x32 env_name=cheetah -ift6163-v0 num_agent_train_steps_per_iter=500 –n_layers 1 size=32

python run_hw2_mb.py exp_name=q1_cheetah_n5_arch2x250 env_name=cheetah-ift6163-v0 num_agent_train_steps_per_iter=5 n_layers=2 size=250

python run_hw2_mb.py exp_name=q1_cheetah_n500_arch2x250 env_name= cheetah-ift6163-v0 num_agent_train_steps_per_iter=500 n_layers=2

size=250

Your code will produce plots inside your logdir that illustrate your model pre-diction error (MPE) (see ift6163.infrastructure.logger for details¹). The code will also produce a plot of the losses over time. For the first com-mand, the loss should go below 0.2 by the iteration 500. These plots illustrate, for a fixed action sequence, the difference between your model’s predictions (red) and the ground-truth states (green). Each plot corresponds to a differ-ent state element, and the title reports the mean mean-squared-error across all state elements. As illustrated in the commands above, try different neural network architectures as well different amounts of training. Compare the results by looking at the loss values (i.e., itr 0 losses.png), the qualitative model predictions (i.e., itr 0 predictions.png), as well as the quantitative MPE values (i.e., in the title of itr 0 predictions.png).

• For those that want to learn more about logging you should look into commet.ml or weights and biases. I highly recommend you learn about this great tools that make accessing and sharing your experiments much easier.

What to submit: For this question, submit the qualitative model predictions (itr 0 predictions.png) for each of the three runs above. Comment on which model performs the best and why you think this might be the case.

Note that for these qualitative model prediction plots, we intend for you to just copy the png images produced by the code.

Problem 2

What will you implement:

Action selection using your learned dynamics model and a given reward func-tion.

What code files to fill in:

1. ift6131/policies/MPC_policy.py (all lines labeled TODO(Q2), i.e. everything except the CEM section)

What commands to run:

python run_hw2_mb.py exp_name=q2_obstacles_singleiteration env_name= obstacles-ift6163-v0 num_agent_train_steps_per_iter=20

batch_size_initial=5000 batch_size=1000 mpc_horizon=10 video_log_freq=-1

Recall the overall flow of our rl trainer.py. We first collect data with our policy (which starts as random), we then train our model on that collected data, and we then evaluate the resulting MPC policy (which now uses the trained model). To verify that your MPC is indeed doing reasonable action selection, run the command above and compare Train AverageReturn (which was the execution of random actions) to Eval AverageReturn (which was the execution of MPC using a model that was trained on the randomly collected training data). You can expect Train AverageReturn to be around −160 and Eval AverageReturn to be around −70 to −50.

What to submit:

Submit this run as part of your run logs, and include a plot of Train AverageReturn and Eval AverageReturn in your pdf. Note that these will just be single dots on the plot, since we ran this for just 1 iteration.

Problem 3

What will you implement:

MBRL algorithm with on-policy data collection and iterative model training.

What code files to fill in:

None. You should already have done everything that you need, because rl trainer.py already aggregates your collected data into a replay buffer. Thus, iterative train-ing means to just train on our growing replay buffer while collecting new data at each iteration using the most newly trained model.

What commands to run:

python run_hw2_mb.py exp_name=q3_obstacles env_name=obstacles-ift6163-v0 num_agent_train_steps_per_iter=20 batch_size_initial=5000 batch_size=1000 mpc_horizon=10 n_iter=12 video_log_freq=-1

python run_hw2_mb.py exp_name=q3_reacher env_name=reacher-ift6163-v0

mpc_horizon=10 num_agent_train_steps_per_iter=1000

batch_size_initial=5000 batch_size=5000 n_iter=15 video_log_freq=-1

python run_hw2_mb.py exp_name=q3_cheetah env_name=cheetah-ift6163-v0

mpc_horizon=15 num_agent_train_steps_per_iter=1500

batch_size_initial=5000 batch_size=5000 n_iter=20 video_log_freq=-1

You should expect rewards of around −25 to −20 for the obstacles env (takes

∼ 40 minutes), rewards of around −250 to −300 for the reacher env (takes 2 −3 hours), and rewards of around 250 − 350 for the cheetah env takes 3 − 4 hours. All numbers assume no GPU.

What to submit:

Submit these runs as part of your run logs, and include the performance plots in your pdf.

Problem 4

What will you implement:

You will compare the performance of your MBRL algorithm as a function of three hyperparameters: the number of models in your ensemble, the number of random action sequences considered during each action selection, and the MPC planning horizon.

What code files to fill in:

None.

What commands to run:

python run_hw2_mb.py exp_name=q4_reacher_horizon5 env_name=reacher-

ift6163-v0 –add_sl_noise mpc_horizon=5

mpc_action_sampling_strategy=’random’

num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15

video_log_freq=-1 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_horizon15 env_name=reacher-

ift6163-v0 –add_sl_noise mpc_horizon=15

num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15

video_log_freq=-1 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_horizon30 env_name=reacher-

ift6163-v0 –add_sl_noise mpc_horizon=30

num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15

video_log_freq=-1 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_numseq100 env_name=reacher-

ift6163-v0 –add_sl_noise mpc_horizon=10

num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15 —

mpc_num_action_sequences 100 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_numseq1000 env_name=reacher-

ift6163-v0 –add_sl_noise mpc_horizon=10

num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15

video_log_freq=-1 –mpc_num_action_sequences 1000

mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_ensemble1 env_name=reacher-ift6163-v0 –ensemble_size 1 –add_sl_noise mpc_horizon=10 num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15 video_log_freq=-1 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_ensemble3 env_name=reacher-ift6163-v0 –ensemble_size 3 –add_sl_noise mpc_horizon=10 num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15 video_log_freq=-1 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q4_reacher_ensemble5 env_name=reacher-ift6163-v0 –ensemble_size 5 –add_sl_noise mpc_horizon=10 num_agent_train_steps_per_iter=1000 batch_size=800 n_iter=15 video_log_freq=-1 mpc_action_sampling_strategy=’random’

What to submit:

1. Submit these runs as part of your run logs.

2. Include the following plots (as well as captions that describe your observed trends) of the following:

1. • effect of ensemble size

1. • effect of the number of candidate action sequences

1. • efffect of planning horizon

Be sure to include titles and legends on all of your plots, and be sure to gener-ate your plots by extracting the corresponding performance numbers from your saved tensorboard eventfiles.

Problem 5

What will you implement:

You will compare the performance of your MBRL algorithm with action select-ing performed by random-shooting (what you have done up to this point) and CEM.

Because CEM can be much slower than random-shooting, we will only run MBRL for 5 iterations for this problem. We will try two hyperparameter settings for CEM and compare their performance to random-shooting.

What code files to fill in:

1. ift6131/policies/MPC_policy.py

What commands to run:

python run_hw2_mb.py exp_name=q5_cheetah_random env_name=’cheetah-ift6163-v0′ mpc_horizon=15 num_agent_train_steps_per_iter=1500 batch_size_initial=5000 batch_size=5000 n_iter=5 video_log_freq=-1 mpc_action_sampling_strategy=’random’

python run_hw2_mb.py exp_name=q5_cheetah_cem_2 env_name=’cheetah-

ift6163-v0′ mpc_horizon=15 –add_sl_noise

num_agent_train_steps_per_iter=1500 batch_size_initial=5000

batch_size=5000 n_iter=5 video_log_freq=-1

mpc_action_sampling_strategy=’cem’ cem_iterations=2

python run_hw2_mb.py exp_name=q5_cheetah_cem_4 env_name=’cheetah-

ift6163-v0′ mpc_horizon=15 –add_sl_noise

num_agent_train_steps_per_iter=1500 batch_size_initial=5000

batch_size=5000 n_iter=5 video_log_freq=-1

mpc_action_sampling_strategy=’cem’ cem_iterations=4

You should expect rewards of 800 or higher when using CEM on the cheetah env. The final CEM run takes 2 − 3 hours on GPU, and over twice as long without GPU, so we recommend getting started early and using a GPU is possible!

What to submit:

1. Submit these runs as part of your run logs.

2. Include a plot comparing random-shooting with CEM, as well as captions that describe how CEM affects results for different numbers of sampling itera-tions (2 vs. 4).

Be sure to include a title and legend on your plot, and be sure to generate your plot by extracting the corresponding performance numbers from your saved ten-sorboard eventfiles.

Submission

3.2 Submitting the PDF

Your report should be a PDF document containing the plots and responses indicated in the questions above.

3.3 Submitting the Code and Logs

In order to turn in your code and experiment logs, create a folder that contains the following:

• A folder named data with all the experiment runs from this assignment. Do not change the names originally assigned to the folders, as specified by exp name in the instructions. To minimize submis-sions size, please include runs with video logging disabled. If you would like to reuse your video logging runs, please see the script provided in ift6163/scripts/filter_events.py.

• The ift6131 folder with all the .py files, with the same names and directory structure as the original homework repository (not include the data/ folder). A plotting script should also be submitted, which should be a python script (or jupyter notebook) such that running it can generate all plots from your pdf. This plotting script should extract its values directly from the experiments in your run logs and should not have hardcoded reward values.

As an example, the unzipped version of your submission should result in the following file structure. Make sure that the submit.zip file is below 15MB and that they include the prefix hw2 mb .

submit.zip

data

hw2 q1 …

events.out.tfevents.1567529456.e3a096ac8ff4 …

ift6131

agents

mb agent.py

…

policies

…

…

README.md

…

If you are a Mac user, do not use the default “Compress” option to create the zip. It creates artifacts that the autograder does not like. You may use zip -vr submit.zip submit -x “_*.DS Store” from your terminal.

Turn in your assignment on Gradescope. Upload the zip file with your code and log files toHW2 Code, and upload the PDF of your report to HW2.

13