tf-mdp 0.5.0

Probabilistic planning in continuous state-action MDPs using TensorFlow.

# tf-mdp [![Build Status][travis.svg]][travis] [![Documentation Status][readthedocs]][readthedocs-badge] [![License][license.svg]][license]

Probabilistic planning in continuous state-action MDPs using TensorFlow.

**tf-mdp** is an implementation based on the paper:

> Thiago P. Bueno; Leliane N. de Barros; Denis D. Mauá; Scott Sanner<br>
> **Deep Reactive Policies for Planning in Stochastic Nonlinear Domains**<br>
> In AAAI, 2019.

# Quickstart

**tf-mdp** is a Python3.5+ package available in PyPI.

```text
$ pip3 install tf-mdp
```

Please make sure you have a running TensorFlow version on your system before pip-installing this package.

# Features

**tf-mdp** solves discrete time MDPs with continuous state-action spaces.

The domains/instances are specified using the [RDDL][rddl] language.

It is built on the following packages available on the Python3 RDDL toolkit:

- [pyrddl][pyrddl]: RDDL lexer/parser.
- [rddlgym][rddlgym]: A toolkit for working with RDDL domains.
- [rddl2tf][rddl2tf]: RDDL2TensorFlow compiler.
- [tf-rddlsim][tf-rddlsim]: A RDDL simulator running in TensorFlow.

Please refer to each project documentation for further details.


# Usage

```text
$ tfmdp --help
usage: tfmdp [-h] [-l LAYERS [LAYERS ...]]
[-a {none,sigmoid,tanh,relu,relu6,crelu,elu,selu,softplus,softsign}]
[-iln] [-hln] [-kl1 KERNEL_L1_REGULARIZER]
[-kl2 KERNEL_L2_REGULARIZER] [-bl1 BIAS_L1_REGULARIZER]
[-bl2 BIAS_L2_REGULARIZER] [-est {pd,sf,hybrid}] [-n N_STEP]
[-bl] [-b BATCH_SIZE] [-hr HORIZON] [-e EPOCHS]
[-lr LEARNING_RATE]
[-opt {Adadelta,Adagrad,Adam,GradientDescent,ProximalGradientDescent,ProximalAdagrad,RMSProp}]
[-lfn {linear,mse}] [-ld LOGDIR] [-v]
rddl

Probabilistic planning in continuous state-action MDPs using TensorFlow.

positional arguments:
rddl RDDL file or rddlgym domain id

optional arguments:
-h, --help show this help message and exit
-l LAYERS [LAYERS ...], --layers LAYERS [LAYERS ...]
number of units in each hidden layer in policy network
-a {none,sigmoid,tanh,relu,relu6,crelu,elu,selu,softplus,softsign}, --activation {none,sigmoid,tanh,relu,relu6,crelu,elu,selu,softplus,softsign}
activation function for hidden layers in policy
network
-iln, --input-layer-norm
input layer normalization flag
-hln, --hidden-layer-norm
hidden layer normalization flag
-kl1 KERNEL_L1_REGULARIZER, --kernel-l1-regularizer KERNEL_L1_REGULARIZER
kernel L1-regularizer constant (default=0.0)
-kl2 KERNEL_L2_REGULARIZER, --kernel-l2-regularizer KERNEL_L2_REGULARIZER
kernel L2-regularizer constant (default=0.0)
-bl1 BIAS_L1_REGULARIZER, --bias-l1-regularizer BIAS_L1_REGULARIZER
bias L1-regularizer constant (default=0.0)
-bl2 BIAS_L2_REGULARIZER, --bias-l2-regularizer BIAS_L2_REGULARIZER
bias L2-regularizer constant (default=0.0)
-est {pd,sf,hybrid}, --estimator {pd,sf,hybrid}
gradient estimator (pd=pathwise derivatives, sf=score
function, hybrid=pd+sf) (default=pd)
-n N_STEP, --n-step N_STEP
n-step hybrid estimator
-bl, --baseline baseline flag
-b BATCH_SIZE, --batch-size BATCH_SIZE
number of trajectories in a batch (default=256)
-hr HORIZON, --horizon HORIZON
number of timesteps (default=40)
-e EPOCHS, --epochs EPOCHS
number of timesteps (default=200)
-lr LEARNING_RATE, --learning-rate LEARNING_RATE
optimizer learning rate (default=0.001)
-opt {Adadelta,Adagrad,Adam,GradientDescent,ProximalGradientDescent,ProximalAdagrad,RMSProp}, --optimizer {Adadelta,Adagrad,Adam,GradientDescent,ProximalGradientDescent,ProximalAdagrad,RMSProp}
loss optimizer (default=RMSProp)
-lfn {linear,mse}, --loss-fn {linear,mse}
loss function (default=linear)
-ld LOGDIR, --logdir LOGDIR
log directory for data summaries (default=/tmp/tfmdp)
-v, --verbose verbosity mode
```

# Examples

```text
$ tfmdp Reservoir-20 -l 2048 -iln -a elu -b 256 -hr 40 -e 200 -lr 0.001 -lfn mse -v

Running tf-mdp v0.4.4 ...

>> RDDL: Reservoir-20
>> logdir: /tmp/tfmdp

>> Policy Net:
layers = [2048]
activation = elu
input layer norm = True
hidden layer norm = False

>> Hyperparameters:
epochs = 200
learning rate = 0.001
batch size = 256
horizon = 40

>> Optimization:
estimator = pd
optimizer = RMSProp
loss function = mse
kernel l1-regularization = 0.0
kernel l2-regularization = 0.0
bias l1-regularization = 0.0
bias l2-regularization = 0.0

>> Loading model ...
Done in 0.058076 sec.

>> Optimizing...
Epoch 199: loss = 1183033344.000000
Done in 251.438551 sec.

>> Performance:
total reward = -4019.2837, reward per timestep = -100.4821

tensorboard --logdir /tmp/tfmdp/train
```

```text
$ tfmdp HVAC-3 -l 256 128 64 32 -iln -a elu -b 256 -hr 40 -e 200 -lr 0.0001 -lfn mse -v

Running tf-mdp v0.4.4 ...

>> RDDL: HVAC-3
>> logdir: /tmp/tfmdp

>> Policy Net:
layers = [256,128,64,32]
activation = elu
input layer norm = True
hidden layer norm = False

>> Hyperparameters:
epochs = 200
learning rate = 0.0001
batch size = 256
horizon = 40

>> Optimization:
estimator = pd
optimizer = RMSProp
loss function = mse
kernel l1-regularization = 0.0
kernel l2-regularization = 0.0
bias l1-regularization = 0.0
bias l2-regularization = 0.0

>> Loading model ...
Done in 0.045406 sec.

>> Optimizing...
2019-03-23 15:57:55.761501: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2
Epoch 199: loss = 24277512192.0000000
Done in 93.795296 sec.

>> Performance:
total reward = -306241.7500, reward per timestep = -7656.0437

tensorboard --logdir /tmp/tfmdp/train
```

```text
$ tfmdp Navigation-v2 -l 256 128 64 32 -a elu -b 128 -hr 20 -e 200 -lr 0.001 -lfn mse -v

Running tf-mdp v0.4.4 ...

>> RDDL: Navigation-v2
>> logdir: /tmp/tfmdp

>> Policy Net:
layers = [256,128,64,32]
activation = elu
input layer norm = False
hidden layer norm = False

>> Hyperparameters:
epochs = 200
learning rate = 0.001
batch size = 128
horizon = 20

>> Optimization:
estimator = pd
optimizer = RMSProp
loss function = mse
kernel l1-regularization = 0.0
kernel l2-regularization = 0.0
bias l1-regularization = 0.0
bias l2-regularization = 0.0

>> Loading model ...
Done in 0.039008 sec.

>> Optimizing...
2019-03-23 16:02:54.502184: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2
Epoch 199: loss = 555.7142338
Done in 28.249370 sec.

>> Performance:
total reward = -79.6495, reward per timestep = -3.9825

tensorboard --logdir /tmp/tfmdp/train
```

# Documentation

Please refer to [https://tf-mdp.readthedocs.io/][readthedocs] for the code documentation.


# Support

If you are having issues with tf-mdp, please let me know at: [[e-mail]](mailto://[e-mail]).

# License

Copyright (c) 2018-2019 Thiago Pereira Bueno All Rights Reserved.

tf-mdp is free software: you can redistribute it and/or modify it
under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation, either version 3 of the License, or (at
your option) any later version.

tf-mdp is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser
General Public License for more details.

You should have received a copy of the GNU Lesser General Public License
along with tf-mdp. If not, see http://www.gnu.org/licenses/.

[license.svg]: https://img.shields.io/aur/license/yaourt.svg
[license]: https://github.com/thiagopbueno/tf-mdp/blob/master/LICENSE
[pyrddl]: https://github.com/thiagopbueno/pyrddl
[rddl2tf]: https://github.com/thiagopbueno/rddl2tf
[rddl]: http://users.cecs.anu.edu.au/~ssanner/IPPC_2011/RDDL.pdf
[rddlgym]: https://github.com/thiagopbueno/rddlgym
[readthedocs-badge]: https://tf-mdp.readthedocs.io/en/latest/?badge=latest
[readthedocs]: https://readthedocs.org/projects/tf-mdp/badge/?version=latest
[tf-rddlsim]: https://github.com/thiagopbueno/tf-rddlsim
[travis.svg]: https://travis-ci.org/thiagopbueno/tf-mdp.svg?branch=master
[travis]: https://travis-ci.org/thiagopbueno/tf-mdp