AIXIjs

AIXIjs is a JavaScript demo for running General Reinforcement Learning (RL) agents in the browser. In particular, it provides a general and extensible framework for running experiments on Bayesian RL agents in general (partially observable, non-Markov, non-ergodic) environments.

UPDATE (May 2017): I'll be presenting a conference paper containing a literature survey along with some experiments based on AIXIjs at IJCAI 2017, in Melbourne, Australia. The paper (to appear) is: J. S. Aslanides, Jan Leike, and Marcus Hutter. "Universal Reinforcement Learning Algorithms: Survey & Experiments", in Proceedings of the 26th Intl. Joint Conf. on A.I..

Features

Agents:

• Bayes (AIXI)
• KSA (Square, Shannon, and Kullback-Leibler)
• Thompson Sampling
• MDL Agent
• BayesExp
• (Tabular) QLearning/SARSA

Environments:

• Bandits
• Finite-state MDPs
• Gridworld POMDPs

See the main site for more background, documentation, references, and demos.

Note: AIXIjs uses some features from ECMAScript 2015 (ES6). It should work on recent versions of Firefox, Safari, Edge, and Chrome, but it's only really been tested on Chrome, so that's what I recommend using.

API

There are a number of demos pre-made and ready to go; look at src/demo.js for examples. If you want to roll your own demo, here's an example of how to get a basic simulation working, with AIXI on a Dispenser Gridworld:

let config = { /* ... */ }; // environment config; see src/config.js for examples
let env = new Gridworld(config); // construct environment
let options = { /* ... */ }; // agent options; see src/config.js for examples
let agent = new BayesAgent(options); // construct agent
let trace = new BayesTrace(); // accumulator for performance and history

let a = null; // action
let e = env.generatePercept() // percept

// main loop
for (let t = 0; t < options.cycles; t+s+) {
	trace.log(agent, env, a, e); // log info to trace
	a = agent.selectAction(e); // agent computes its policy and selects an action
	env.perform(a); // pass this action to the environment and compute dynamics
	e = env.generatePercept(); // environment emits a percept
	agent.update(a, e); // agent learns from action and resulting percept
}
// now do what you want with the trace -- it has logged all the relevant data
//

Note that agents should implement two methods, selectAction(e) and update(a,e). Environments should implement generatePercept(), perform(a), and conditionalDistribution(e).

How to run experiments

I've provided a helper function demo.experiment(...configs,runs) to make it easy to do numerous runs with different configs, and serialize the results to JSON. To run experiments, bring up the console in Chrome (Ctrl+Shift+I on Linux), and run something like:

After the experiments download link for results.json will appear at the bottom of the DOM. See src/demo.js:209 for details about the structure of the serialization. There's an iPython notebook in experiments/analysis.ipynb that should help get you started in producing plots etc.

License

GPL.

How to cite

If you use this software in your own experiments, please cite it as:

@inproceedings{ALH2017,
 author = {Aslanides, John and Leike, Jan and Hutter, Marcus},
 title = {Universal Reinforcement Learning Algorithms: Survey and Experiments},
 booktitle = {Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence},
 series = {IJCAI'17},
 year = {2017},
 publisher = {AAAI Press},
}

Contributing

If you'd like to contribute, I'm all ears! There's a lot of stuff to do :) Go ahead and open a pull request!