Home | Products | Downloads | Research | Contacts | Forum | Links

Tantalus Machine Learning System

    For those interested, we have released a demo version of our Tantalus machine learning software. Tantalus is a multiple competing agent architecture, utilizing 6 standard (4 in the demo) agents, each applying a different learning technique. 

    For the demonstration we packaged the learning system in a conversational training application. This is basically an example of using machine learning to teach the application conversational skills. Tantalus was a product of our research prior to our Xanthus and Hermes research projects, and has been shelved for the last year or so. 

    We thought it would be interesting to release the demonstration for comment. So please feel free to comment on our website forum. 

    Tantalus has been set aside for more promising technologies (like our Xanthus and Hermes research), but it isn't a "failure" so much as conceptually inadequate for true volitional intelligent systems. It is more like an idiot savant. 

The 6 agents in the Tantalus system are: 

1 - GA and Probabilistic 
2 - Fuzzy operant conditioning 
3 - Temporal fuzzy operant conditioning 
4 - A-Level inference system 
5 - B-Level inference system 
6 - Categorical and relational 

    In the demo only 1 - 4 are implemented. The Tantalus system was originally developed for real-time embodied robotic experiments, so porting to an NLP application has been a bit of work.  We will finish the job if there is sufficient demand from users.

Using the demo program:

    First, please note that the demonstration is not intended to be a chatbot!  The purpose of the demonstration software is to illustrate the efficacy of the Tantalus machine learning system when applied to a specific task, which in this case is learning conversational skills.  The demo therefore does not come pre-trained, but instead comes entirely bereft of knowledge.  How quickly you can take this tabula rasa system and teach it to converse is the point of the demonstration.

    To begin using the Tantalus demo software simply type a sentence in the "Sentence In" edit box and press "Best Response".  The first time you do this the response will not be very good, because Tantalus doesn't know anything at all yet (not a single word!).  After pressing the button you will note that the words you typed will appear in the "Dictionary" list.  These are words that Tantalus will recognize in the future.

    To start training Tantalus you must rate the response received.  This is done in the rating and alternate response section beneath the sentence in and response area.  The system does not remember anything unless it is explicitly told to "Add To History".  The rating that you specify refers to the response that was provided to the last sentence in.  Simply pick a rating from the scale.  This is how you provide Tantalus with positive and negative reinforcement for it's behavior.

    A powerful training tool is provided by the use of the "Alternate Response" input.  This is how you give Tantalus an example of a Good response.  The learning and inference agents will then use these response in an attempt to formulate future responses to input.  By reinforcing positively or negatively the system will get better at conversing.

    The "Think" button is used to activate the higher inference learning behavior.  This is done explicitly in the demo, whereas in a real application it would constantly run in a background thread.  Once you have added a number of trainings to the history you should click the "Think" button to allow Tantalus to infer new rules of behavior.  You can abort long thinking sessions with the "Stop" button.

When run the demonstration program will create 3 files in the default location:

tantalus.dic - The dictionary file

tantalus.hst - The history file

tantalus.ars - The A-level inference rules

If you wish to start with a blank system again, simply delete these three files.

With future releases that implement additional agents there may be additional files created.

Example training session:

U: Hello there Tantalus

T: there there tantalus

This is not a great response, so Rating is set to "Poor", and Alternate Response is set to: Hello human, click "Add To History"

U: Hello there Tantalus

T: hello human

U: Time to say goodbye

T: goodbye goodbye say to to tantalus tantalus hello time there goodbye

That isn't right, so Rate as Poor, and alternate to: Goodbye, add to history

U: Time to say goodbye

T: goodbye 

U: I like rootbeer

T: i there time

That isn't right, so Rate as Poor, and set alternate to: Humans like rootbeer, add to history

U: I like chocolate

T: humans like rootbeer

That isn't right, so Rate as Poor, and set alternate to: Humans like chocolate, add to history

Now click "Think" so the inference agents can come up with new rules

U: I like bananas

T: humans like bananas

Tantalus has learned a new rule!  So rate this as Good and add to history.

U: I like ice-cream

T: humans like ice-cream

The system has now mastered this rule.  There is no literal limit to the number of variables that can be inferred in such fashion.

Please post any questions or comments in our online Forum.  Thank you!

Return to Research

Home