The reinforcement learning paradigm has the aim to determine the optimal actions for a robot. A typical example is a maze finding robot, but reinforcement learning can also be used for training a robot to play the pong game. The principle is based on a reward function. If the robot is able to solve a problem, he gets a score from the underlying game engine. The score can be positive, if the robot reaches the end of a maze, or it can be negative, if he is colliding with an obstacle. The principle itself is working quite well, that means for simpler applications it is possible to train a robot to play a game with reinforcement learning.

Chatbots are a different category of artificial intelligence. They are working not with actions but with natural language. Person #1 is opening a dialogue with “Hi, I'm Alice”, while person #2 is responding with “Nice to meet you”. What is missing here is an underlying game which is played. There is no reward available for printing out a certain sentence. In some literature the problem of language grounding was discussed seriously, but with an unclear result. It seems, that a classical game for example pong, and a chatbot conversation doesn't have much in common.

Is it possible to combine Reinforcement Learning with chatbot design? The problem is, that a speech-act should be connected to a reward. That means, a well formulated sentence gets +10 points but a weak sentence gets -10 points. How can this be evaluated?


The problem is indeed that the 'rules' of conversations are not as fixed as the rules for games. However, you can make use of descriptive formalism from Discourse Analysis, called adjacency pairs. These describe regularities between utterances on a local level, for example greeting/reply, which would match your "Hi, I'm Alice" and "Nice to meet you".

You will need to be able to classify utterances by your chat bot according to a set of possible responses, and then you can see if a valid response is produced for any given utterance. If the user asks a question, then a greeting will not be a good answer, but a statement could be, if it was a response to the question. This is leaving aside the content and focuses merely on the formal characteristics of the utterance.

If you want to know more about the topic, have a look at Conversation Analysis, which is the linguistic field dealing with the subject.

| improve this answer | |
  • $\begingroup$ This makes me think it might be worthwhile to formally explore games with fluid rulesets. (Games, in general, evolve over time, but I wonder if there are any combinatorial games in which rules emerge from gameplay.) $\endgroup$ – DukeZhou Mar 1 '19 at 22:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.