Files
2019-06-29 02:43:12 +02:00

122 lines
4.1 KiB
ReStructuredText

Rewards
=======
In PRL, every concept is modelized as a class. This is also true for rewards which are returned by the environment as shown in the figure below (inspired by [1]_):
.. figure:: ../figures/environment.png
:alt: environment
:align: center
The agent-environment interaction
The reward function might be defined as [1]_:
- :math:`r: \mathcal{S} \rightarrow \mathbb{R}`: given the state :math:`s \in \mathcal{S}`, it returns the reward value :math:`r(s)`.
- :math:`r: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}`: given the state :math:`s \in \mathcal{S}` and action :math:`a \in \mathcal{A}`, it returns the reward value :math:`r(s,a)`.
- :math:`r: \mathcal{S} \times \mathcal{A} \times \mathcal{S} \rightarrow \mathbb{R}`: given the state :math:`s \in \mathcal{S}`, action :math:`a \in \mathcal{A}`, and next state :math:`s' \in \mathcal{S}`, it returns the reward value :math:`r(s,a,s')`.
Note that the cost function is just minus the reward function, i.e. it is given by :math:`c(s,a,s') = -r(s,a,s')`.
Design
------
In PRL, all reward functions inherit from the abstract ``Reward`` class defined in `pyrobolearn/rewards/reward.py <https://github.com/robotlearn/pyrobolearn/tree/master/pyrobolearn/rewards>`_, and several methods and operations are provided.
You can for instance:
* provide the ``State`` and/or ``Action`` instances to some reward functions that will compute the reward value based on their value.
* access to the range of the reward function.
* add, multiply, divide, subtract, and apply basic functions such as :math:`\exp`, :math:`\cos`, :math:`\sin`, and others on reward functions. The resulting range is automatically scaled based on the operations.
* define your own rewards/costs and reuse them in your code.
How to use a particular reward?
-------------------------------
Here is a short snippet showing the basic usage of reward functions:
.. code-block:: python
:linenos:
import pyrobolearn as prl
from pyrobolearn.rewards import FixedReward, YourReward
# define the simulator and world (and load what you want in it)
sim = ...
world = ...
...
# define your state / action for your reward function
state = ...
action = ...
# define the reward function
reward = 2 * FixedReward(3) + 0.5 * YourReward(state, action)
# print the range of the reward function
print(reward.range)
# compute the reward value
value = reward()
print(value)
# update the state for instance
state() # this will modify the internal state data
# recompute the reward value
value = reward()
print(value) # you will normally get a different value
# you can give the reward function to your RL environment
# which will use it when calling `env.step()`.
env = prl.envs.Env(world, state, reward)
More examples on how to use the rewards can be found in `pyrobolearn/examples/rewards <https://github.com/robotlearn/pyrobolearn/tree/master/examples/rewards/>`_.
How to create your own reward?
------------------------------
In order to create your own reward, you have to inherit from ``Reward`` defined in `pyrobolearn/rewards/reward.py <https://github.com/robotlearn/pyrobolearn/tree/master/pyrobolearn/rewards>`_.
.. code-block:: python
:linenos:
import pyrobolearn as prl
class MyReward(prl.rewards.Reward):
"""Description"""
def __init__(self, args):
# initialize your reward function based on the args
...
# gives initial value to your reward
# this attribute will be used to cache the computed value
self.value = 0
def _compute(self):
# compute the reward function
...
# save the computed value and return it
self.value = ...
return self.value
Once done, you will be able to use your reward function and perform operations on it (such as addition, substraction, etc).
FAQs
----
* If you have any questions, please submit an issue on the `Github page <https://github.com/robotlearn/pyrobolearn>`_.
References:
-----------
.. [1] "Reinforcement Learning: An Introduction", Sutton and Barto, 1998