Simple Markov Decision Process Library
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
src
.gitignore
LICENSE
README.md

README.md

Markov Decision Process Rendering Library and Game for Javascript and the Web

A Markov Decision Process is a schocastic method encodes action policies from s to s'. I'll have a blog post when I'm done with this library on this.

Libaries

  • jquery //used to interact with the dom
  • math.js //used for matrix operations.
  • d3 //used for rendering and shape help.
  • node.js / express - to run a server and see examples.

Models

Basic MDP

  • S: finite set of states
  • A: finite set of actions
  • P: SxA -> \Delta{S} A Markovian Transition Function
  • P(s'|s,a) -> The probability of going to s' from s given action a.
  • R: AxS -> |R: Reward function
  • R(a, s') Reward abtained when a moves to s'

POMDP Add an obsevation function and observation likeihood function

  • \Omega: a finite set of observations
  • O: An observation function
  • O(o|a,s'): Probability of observing o when action a is taken to s'

Dec-POMDP and Dec-POMDP-Com This library will eventually include examples for Dec-POMDP and Dec-POMDP-COM.

Policy Optimization

  1. Value Iteration
  2. Q Learning

Transition Approximation

There are three methods for transition approximation that will originally be offered.

  1. Crawl (Agent will randomly crawl and generate probability matrix)
  2. MCMC (Monte Carlo Simulation)
  3. Tree Pruning

To Run

You can download the src file with

   cd examples/simplegame/
   node serve.js

Examples

  • Grid Examples