easyagents 1.4.1

reinforcement learning for practitioners.

Reinforcement Learning for Practitioners (v1.4.1, 20Q1)

Status: under active development, breaking changes may occur. Release notes.

EasyAgents is a high level reinforcement learning api focusing on ease of use and simplicity. Written in Python and running on top of established reinforcement learning libraries like tf-Agents, tensorforce or keras-rl. Environments are implemented in OpenAI gym. For an example of an industrial application of reinforcement learning see here.

In collaboration with Oliver Zeigermann.

Features

---

• provides the same, simple api across all libraries. Thus you can easily switch between different implementations and you don't have to learn for each of them a new api.
• to create and run any algorithm you only need 2 lines of code, all the parameters are named consistently across all algorithms.
• supports a broad set of different algorithms
• runs inside jupyter notebooks as well as stand-alone, easy to install requiring only a single 'pip install easyagents'.
• easy to understand, ready-made plots and logs to investigate the algorithms and environments behaviour

Note: keras-rl backend is suspended until support for tensorflow 2.0 is available.

Examples

---

from easyagents.agents import PpoAgent
from easyagents.callbacks import plot

ppoAgent = PpoAgent('CartPole-v0')
ppoAgent.train([plot.State(), plot.Loss(), plot.Rewards()])

More Detailed

from easyagents.agents import PpoAgent
from easyagents.callbacks import plot

ppoAgent = PpoAgent( 'Orso-v1',fc_layers=(500,500,500))
ppoAgent.train([plot.State(), plot.Loss(), plot.Rewards(), plot.Actions(), 
                plot.StepRewards(), plot.Steps(), plot.ToMovie()], 
                learning_rate = 0.0001, num_iterations = 500, max_steps_per_episode=50 )

Tutorials

---

• 1. Introduction (CartPole on colab): training, plotting, switching algorithms & backends. Based on the classic reinforcement learning example balancing a stick on a cart.
• 2. Next steps & backend switching (Orso on colab): custom training, creating a movie & switching backends. gym environment based on a routing problem.
• 3. Controlling training & evaluation (on colab): or 'what do all these agent.train(...) args mean ?'
• 4. Creating your own environment (LineWorld on colab): implement a gym environment from scratch, workshop example.
• 5. Saving & loading (on colab): Once a policy is trained, save it and reload it in a production environment. You may also save intermediate policies as the training proceeds.
• 6. Switching backends (on colab): See how you can switch between backend implementations.
• 7. Api logging, seeding & plot clearing (on colab): Investigate how easyagent interacts with the backend api and the gym environment; how to set seeds; controlling jupyter output cell clearing

Available Algorithms and Backends

---

algorithm	tf-Agents	tensorforce	keras-rl (suspended)	easyagents class name
CEM	not available	not available	yes	CemAgent
Dqn	yes	yes	yes	DqnAgent
Double Dqn	open	not available	yes	DoubleDqnAgent
Dueling Dqn	not available	yes	yes	DuelingDqnAgent
Ppo	yes	yes	not available	PpoAgent
Random	yes	yes	not available	RandomAgent
REINFORCE	yes	yes	not available	ReinforceAgent
SAC	preview	not available	not available	SacAgent

[191001]

• if you are interested in other algorithms, backends or hyperparameters let us know by creating an issue. We'll try our best to support you.
• for a documentation of the agents api see here.
• starting with easyagents 1.3 (191102) the backend for keras-rl is suspended until support for tensorflow 2.0 is available.

Industrial Application

---

Geberit - a sanitary technology company with > 12'000 employees - produces in particular pipes and other parts to get rain-water of flat roofs - so called syphonic roof drainage systems. They warrant that large buildings like stadiums, airports or shopping malls do not collapse during heavy rainfalls. However it is surprisingly difficult to find the right dimensions for the pipes. It is actually so difficult, that as of today no feasable, deterministic algorithm is known. Thus traditional heuristics and classic machine learning were used to support the users in finding a suitable solution.

Using reinforcement learning the failrate of the previous solution was reduced by 70%, resulting in an end-to-end success-rate of > 98%.

For more details take a look at this talk.

Installation

---

Install from pypi using pip:

pip install easyagents

More

---

Documentation

for release notes & class diagram, for agents & api.

Guiding Principles

• easily train, evaluate & debug policies for (you own) gym environment over "designing new algorithms"
• simple & consistent over "flexible & powerful"
• inspired by keras:
  • same api across all algorithms
  • support different implementations of the same algorithm
  • extensible (pluggable backends, plots & training schemes)

EasyAgents may not be ideal if

• you would like to leverage implementation specific advantages of an algorithm
• you want to do distributed or in parallel reinforcement learning

Note

• If you have any difficulties in installing or using easyagents please let us know by creating an issue. We'll try our best to help you.
• Any ideas, help, suggestions, comments etc in python / open source development / reinforcement learning / whatever are more than welcome.