figshare
Browse
- No file added yet -

Sources and Data of Submitted Semantic Web Journal Paper:"Context-Aware Composition of Agent Policies by Markov Decision Process Entity Embeddings and Agent Ensembles"

Download (9.57 MB)
journal contribution
posted on 2023-11-07, 16:53 authored by Nicole MerkleNicole Merkle, Ralf Mikut

SW Journal

Here you can find the source code and generated datasets regarding the domestic activities that were generated in the scope of the submitted Semantic Web Journal paper entitled: "Context-Aware Composition of Agent Policies by Markov Decision Process Entity Embeddings and Agent Ensembles". The generated data is based on the Virtual Home dataset, which is published under the MIT license.


Directory Structure

  • code -- Contains the source code
  • Entity_Embeddings -- contains the trained entity embeddings
  • VirtualHomeKG.zip -- contains the generated semantic task descriptions of all activities given in the Virtual Home dataset.

Preparations

The demo was tested on a Windows OS. Other operating systems may require additional libraries.


  • Install Nodejs + NPM on your host system.
  • Download zip file to your host system and unpack it to your local host system.


Configuration of the Demo (optional)

It is not required to change the configuration of the demo. However, you have the possibility to specify the number of tasks to be evaluated as well as the radius hyper-parameter of the executed algorithm. The .env configuration file contains all configurable parameters as key-value pairs. The following parameters to be defined:


  • NUMBER_TASKS -- The number of tasks that shall be evaluated. Default value is set to 52 as performed in the evaluation of the presented approach.
  • RADIUS -- The search radius for finding nearby embedding vectors. Default value is set to 0.25

Start Demo:

Go into the code directory and execute via command line:

npm install


All required javascript libraries are installed automatically.


  • Finally, execute: npm start
  • (Optional) The deep neural network for training the entity embeddings can be executed via: node EmbeddingTrainer.js

It is not required, i.e. optional, to train the entity embeddings from new because they are already available in the Entity_Embeddings directory. On some systems pre-node-gyp issues may occur, since maybe the python interpreter or some compiler versions or libraries may be missing. Please install missing and required software in cases when such issues occur.


Obtained Results

Two files (AgentEnsemblesEval.csv and ComposedPolicies.json) will be generated that contain the evaluation results, such as task name, required steps, number of episodes, cumulative reward, wrong decisions and the policies that were composed for each evaluated task. Furthermore, a directory named ../VirtualHomeKG will be generated that contains two directories (JSONLD and TURTLE). In both directories the semantic task representations for the simulation function are stored in JSONLD and Turtle format.




History