Terminology
Model
user: The agent that is taking part in an interaction with a certain goal. It may be a real human user, or a synthetic user model.
assistant: an intelligent tool or interface that assists the user.
agents: some entity that can produce observations, perform inferences and take actions. Refers in practice to both users and assistants.
interaction task, task: a task defined by a given state, to be solved by the user with the assistant’s aid. A task is solved when it reaches a certain goal state.
interaction game, game: the two-agent game formed by combining a task, a user, and an assistant.
Module
bundle: the component that turns a task, a user and an assistant into an interaction game
observation engine: the agent’s component that produces observations from the game state
inference engine: the agent’s component that is responsible for modifying its internal state, based on the previous state and the available observations.
sampling, sample: selecting an action from the possible action set. Usually this is not random sampling, but based on some policy.
round: Each sequence of joint decisions and task updates make a round (just like in a boardgame, a round is when everyone has played its turn)
turn: There are 4 turns per round (observation and inference, taking action, both for the user and the assistant).
step: round
reset: Force the state of the environment to an initial (potentially random) state.