Skip to content
This repository has been archived by the owner on Mar 31, 2019. It is now read-only.

Whats and Whys

Jump to bottom
justheuristic edited this page May 4, 2016 · 3 revisions

What?

AgentNet is a lightweight library for deep reinforcement learning using Theano+Lasagne. It also provides support for attentive and long-term memory networks.

The main objectives are:

  • Easy prototyping for reinforcement learning architectures
  • Support for arbitrary recurrent, attentive and long-term memory networks
  • Ease of experiment conduction and reproducibility
  • Full integration with Lasagne and Theano
  • [next step - full integration with TensorFlow]

Why?

[long story short: create a platform to play with *QN, attentive and LTM architectures without spending months reading code]

[short story long:

The last several years have marked the rediscovery of neural networks applied to Reinforcement Learning domain. The idea has first been introduced in early 90's [0] or even earlier, but was mostly forgotten soon afterwards.

Years later, these methods were reborn under Deep Learning sauce and popularized by Deepmind [1,2]. Several other researchers have already jumped into the domain with their architectures [3,4] and even dedicated playgrounds [5] to play with them.

The problem is that all these models exist in their own problem setup and implementation bubbles. Simply comparing your new architecture the ones you know requires

  • 10% implementing architecture
  • 20% implementing experiment setup
  • 70% reimplementing all the other network architectures

This process is not only inefficient, but also very unstable, since a single mistake while implementing 'other' architecture can lead to incorrect results.

So here we are, attempting to build yet another bridge between eager researchers [primarily ourselves so far] and deep reinforcement learning.

The key objective is to make it easy to build new architectures and test is against others on a number of problems. The easier it is to reproduce the experiment setup, the simpler it is to architect something new and wonderful, the quicker we get to solutions directly applicable to real world problems.

]

Clone this wiki locally