KorraAI - a probabilistic virtual agent framework

Editor's Notes

• #2: Last updated July 2022.
• #4: For example humans can get bored when prediction matches result in a computer game. Communication with other people can be considered as both “Need” or “Periodical Activity”. The Von Neumann architecture is one that works with exact data and it is the one used in every computer today. It is not designed (from hardware point of view) to store and handle uncertain information. But it can still be done using programming. There are many probabilistic frameworks available.
• #5: One needs very good NLP and Semantic Analysis to make a good ECA. High expectations because movies make AI so human like in both reasoning and how it looks. NLP with performance of 80% is doable, but then it is a long way between 80 and 82%. Lojban is a carefully constructed spoken language. Its syntactic structure and validity of a sentence is unambiguous. Lojban's grammar is based on simple rules, and its linguistic features are inspired by predicate logic.
• #7: Not hard to install and execute – no need to install second party runtimes such as  Java, extra compilers, etc.
• #9: Also: currently only on one programming language C# (which helps development a lot) Probabilistic Programming and Bayesian Networks are used for modelling
• #10: It can employ different strategies and behaviors over time.
• #12: KorraAI.dll is the plugin and can host several ECAs.  We exchange models as source code or compiled in KorraAI.dll.
• #13: April is the combination of an everyday companion and sales agent. April is closed source.
• #14: On the left we see the “engine” of KorraAI and on the right components 7, 8, 9, 10, 11 and 12 represent the “model” provided by an ECA designer. The engine executes a loop that is defined by components 1,2,3,4. The “Main Distribution” controls the overall behavior and it is stored in component 11. The engine assures that the interactions are loaded from 10, then it uses the Main Distribution to produce the Interactions Queue 3.
• #16: The KorraAI engine consults the ECA model and executes accordingly. One needs to provide at least “Interactions Categories”, “Intertactions and Timing Distributions - “Interaction Categories” can be “The ECA introduces her/his self”, “The ECA asks the user introduction questions”, “Jokes”, “Suggest movies to watch”, “Motivate the user to do sport”. - An “interaction” is for example “Ask the user hers/his name” from the interaction category “The ECA asks the user introduction questions”. An interaction can also be non-verbal. - Triggers are described later in the presentation. They track user’s responses and elapsed time and depending on the type of trigger they can contain a probabilistic model and they can either update the “Main Distribution” (and thus change the ECA behavior) or add a new interaction in the “Interaction Queue”. - KorraAI promote the use of probabilistic variables and probabilistic models.
• #17: Non-deterministic program – the output is not the same every time it the program is executed with the same input. - “Main Distribution” sets which categories should be used to define the principal behavior of the ECA. It also sets how much each Interaction Category should be used. For example Category 1 can be with 0.7 probability, and Category 2 can be with 0.3 probability and thus we will have much more interactions from Category 1. - Each category can also have a distribution. Be default it is the uniform. The idea is that certain interactions in a category should be used before others (depending on elapsed time for example). For example some type of mildly offensive jokes should be reserved for later in the communication. The So we reduce the chance of using these in the beginning, but it is still possible to have a mildly offensive joke in the beginning. - Time should not be fixed. Fixed time intervals will allow the user to get accustomed to the ECA and know what to expect (at which moment). Timing distributions are for example: “Distribution on pause between two interactions”, “When to smile”, “Pause before reacting to the user’s response”, “How much time before the ECA gives up on waiting for user’s response”. Often the normal distribution is used centering around some value in seconds e.g. mean = 20 seconds and variance = 1.5
• #18: Here three distributions are used to select which phrases the ECA will utter. 1) First a category is selected using a probabilistic sampler. For example “Suggest a Movie”. 2) Then we sample over all the interactions inside the category. We select a movie where some movies could be with higher likeliness to be selected – this can be based on the genre of the movie for example. 3) The interaction itself contains code that generates phrases in real time 3) is explained on the next slide
• #19: The name of the movie is stored in the “movieName” variable. This an actual code from KorraAI when a interaction from the category “Suggest a Movie” has been sampled (selected). The following methods can be used: - KorraModelHelper.GetChance – uniformly select an item from an array - KorraModelHelper.GetChance(3) – return true with probability 33% - KorraModelHelper.GetChance(4) – return true with probability 25% The phrase is built dynamically by selecting a principal phrase with KorraModelHelper.GetChance() and then we have 33% chance to add "Do you feel bored? " if principal phrase 3 is selected. Example movie suggestion. The name of the movie is stored in the “movieName” variable. string text = KorraModelHelper.GetChance(new string[] { "How about this " + type + " " + movieName + "? I recommend it.", //1 "I have a new " + type + " suggestion for you. It is called : " + movieName + ". You should try it.", //2 (KorraModelHelper.GetChance(3) ? "Do you feel bored? " : "") + "I recommend you to watch the " + type + " " + movieName + ".", //3 "Time for something interesting to watch. Try this one: " + movieName + ".", //4 "You are going to like this " + type + ": " + movieName + ". I highly recommend it." //5 });
• #20: There are two possibilities: To use a pre-generated Interaction Queue. This is easier to test and it can be inspected at any time. One can verify if the generated interactions really follow the “Main Distribution” from which they were generated. This is the method used in KorraAI. We can not generate the exact text for all interactions in advance, that is why we have placeholders such as “###place holder for InAGoodMood”. The exact text for the mood interaction is generated just before execution based on the current mood of the ECA. Each interaction is supplied (probabilistically sampled) just before it is needed. This method also has advantages: we do not need to track when to regenerate (re-sample) the “Main Distribution” which is needed when communicating with the user or over time. Also we do not need to use placeholders for certain interactions.
• #21: Both are probabilistic models.
• #22: There is a SpeechAdapdation modeule in KorraAI that handles some of the naturalization tasks. Speech: - At the beginning of the session a “Hello” interaction is inserted which uses a uniform distribution over “Hi”, Hello”, “Hello there”, “Hey” and if this is not the first interaction then it adds the word “again”. If the name of the user is available, the name can also be included in the greeting. - When the ECA asks the user his name then it is better that the ECA also introduce her name in this very moment. In general these are two independent interactions in the Interaction Queue and they can be sampled not one after another. - Sometimes we address people by name. But this should not be all the time or it may become annoying (for some people). That is why there is a distribution that sets that calling by name should be only in 33% of the cases and only for specific interaction categories. Non-verbal: - “Reaction time pauses” follows a different distribution (that produces a shorter time interval) than “New interaction pause”. - When the ECA starts a new speech interaction the ECA’s stays focused on the user using a time interval produced by a Normal distribution (mean = 7 seconds, variance = 1.2). - Voice annotation: pauses, pitch, speed, intensity. SSML - Speech Synthesis Markup Language
• #23: Normalization: the sum of probabilities of all categories is 1. After some of the categories are depleted then for example the music suggestions will increase from 10 per hour to 18 for example. This can be unwanted behavior. The user might not like too many music suggestions. So we are not following the desired behavior. The best solution is to keep the proportions for some categories and let the others compensate after normalization. This must be set in advance by the ECA designer.
• #24: PP = Probabilistic Programming BUGS is a software package for performing Bayesian inference Using Gibbs Sampling
• #25: All 3: Javascript, Scala and C# have elements from functional languages. C#'s LINQ is declarative functional programming. WebPPL uses Javascript as host language - Figaro uses Scala as host language - Infer.NET is with host language C# Exact inference is possible for small probabilistic models otherwise it might take a lot of time. Figaro and Infer.NET are very similar – the probabilistic part is provided as a library instead of language extensions. Book for Figaro: "Practical Probabilistic Programming“: https://www.manning.com/books/practical-probabilistic-programming
• #26: We set the initial preferences of the user in the form of the probabilistic variables InAGoodMood and LikesJoke. They can be updated by using a question that maps the predefined answers to probabilities. Here we able to incorporate an answer with a variable level of confidence into the “Main Distribution” concerning the “Tell Joke” category and thus modify the ECA behavior.
• #27: The syntax here is similar to LINQ in C# and SQL in general. This syntax is actually rather intuitive. We define a case when the user is for example not in a good mood and likes jokes in general to be with probability 0.9. This is because we want to cheer her/him up. Another case is 0.4 and all the others are with probability 0.2. Next we define a sampling method that intuitively does the following: it takes samples from LikesJoke and InAGoodMood and applies probabilistic reasoning using the function that we defined previously TellJokeProb. After that TellJokeRate  is used in the "Main Distribution“ for the category “Tell Jokes”.
• #28: 24h time means: is it morning, time to eat, time to sleep, time to rest, etc. Another example of time MUT is to update the “Main Distribution”, so that Movie Suggestions are increased in the evening (after 19h), because we know that the user is occupied, so during the day we should not temp him/her with many movie suggestions.
• #29: POE – probability of observed evidence A contradiction is calculated and then we need to set a threshold – at which value the evidence for contradiction is sufficient, so that we can trigger a response in the form of surprise.
• #30: Forecasted  Interactions Time (FIT) When constructing the “Main Distribution” we provide how much each category should be presented. This works when we sample, for example, a 100 interactions. But in order to be able to correctly model the Main Distribution, it is necessary to know how much time each interaction takes, so that the Main Distribution can be respected for a specific time interval – for example 1 hour. Pi is controlled by a distribution, so we already know its mean value. Ci also known.
• #31: The hidden states can be used in probabilistic models. For example a good education is a factor that usually means that the user will have a higher purchasing power.