BDI-RL Framework Proof of Concept

This project is a PoC of the framework presented in

From Programming Agents to Educating Agents – A Jason-based Framework for Integrating Learning in the Development of Cognitive Agents [DOI] [pre-print (free access)] [talk slides]

Michael Bosello, Alessandro Ricci.

If you use this repo, please cite our EMAS'19 paper.

@InProceedings{10.1007/978-3-030-51417-4_9,
author="Bosello, Michael
and Ricci, Alessandro",
editor="Dennis, Louise A.
and Bordini, Rafael H.
and Lesp{\'e}rance, Yves",
title="From Programming Agents to Educating Agents -- A Jason-Based Framework for Integrating Learning in the Development of Cognitive Agents",
booktitle="Engineering Multi-Agent Systems",
year="2020",
publisher="Springer International Publishing",
address="Cham",
pages="175--194"
}

---

This is an integration of BDI agents and Reinforcement Learning. It is based on Jason (Actually, it is a JaCaMo project).

The basic idea is that a developer could write some plans and let the agent itself learn other plans and use them in a seamless way. This is not only for a specific ad hoc problem but as a general feature of the agent platform.

In short, the aim of the framework is to enable the developer to define the learning components with high-level abstractions as the BDI ones are. Then, these informations injeced by the developer are used by the agent to learn itself how to fulfill some tasks.

The work of the developer moves from write plans to define a learning phase.

As reference examples, we provide several agents and environments.

• The Gridworld env and the SARSA algorithm are implemented in Java ex novo.
• We implemented a RESTful service in Python that provides the capability of TF-Agents to our framework. TF-Agents is a library that offers the core elements of RL and several RL algorithms, which is built on top of TensorFlow. A Java class that implements the interface requested by our framework consumes the above-mentioned service, completing the bridge between our framework and TF-Agents. We tested two algorithms offered by the library: DQN and REINFORCE.
• We provide two classic control environments, namely CartPole and MountainCar. To obtain an ensemble of test tasks, we implemented a bridge – with a REST service, in the same way as for the algorithms – between the Jason environment and Gym, a suite of RL tasks.

Framework usage

The description of the framework, and how to use it, is extensively explained in the paper (and in the slides).

You can develop Jason agents and environments as usual, and use "soft plans" when necessary (along with the appropriate belief). The framework will handle the rest.

If you want to implement your own RL algorithm, you must implement the interface AlgorithmRL, and add its instance in BeliefBaseRL (line 98 to 106)

Src content

• agt/rl contains the framework:

  execute, expected_return two internal actions used to call softplans and check their expected return

  component/ contains the abstractions of the framework

  beliefbase/ the engine of the framework

• agt/rl/algorithm contains the RL algorithm interface and proposed examples (the tf subdir contains the interface to the REST APIs of python algorithms)

• agt/*.asl the agent examples

• env/ contains environment examples (simulation subdir contains common utilities. gym subdir contains the interface to the REST APIs of python envs)

• python/agt implementation of REST APIs exposing tf-agents RL algorithms

• python/env implementation of REST APIs exposing gym RL envs

Quick start

To run the agent system:

./gradlew run

Python algorithms

If you use an algorithm implemented in python, you must run the python agent server before the agent system

Check the dependencies below

You can run the python service with the apposite gradle task:

./gradlew --stop
./gradlew runPythonAgent

or with the python command (from the directory src/python/agt):

python3 tf_agent_rest.py

Python environments

If you use an environment implemented in python, you must run the python environment server before the agent system

Check the dependencies below

You can run the python service with the apposite gradle task:

./gradlew --stop
./gradlew runPythonEnv

or with the python command (from the directory src/python/env):

python3 tf_env_rest.py

Python dependencies

You must install python 3 (System tested on version 3.7.3) along with the following dependencies:

pip3 install tf-agents-nightly
pip3 install tensorflow==2.0.0-beta1
pip3 install 'gym==0.10.11'
pip3 install IPython
pip3 install flask flask-jsonpify flask-restful

Configuration

Algorithm Parameters

You can change the parameters and the rl algorithm (dqn, sarsa, reinforce) with the agent beliefs.

Serialization

From agt/rl.algorithm.BehaviourSerializer:

To serialize learning progress to a file: saveProgress = true

To load learning progress from a file: loadProgress = true;

Gridworld configuration

Show/hide the Gridworld GUI with the SHOW_VIEW boolean in env/gridworld.GridWorldEnv

Save simulation results

To log errors did in each episodes to a file change the boolean SAVE_RESULT in env/simulation.EpisodicSimulation