greedy.py

Go to the documentation of this file.

from agent import BaseAgent
 
 
class Greedy(BaseAgent):
    """
    An agent that always exploits, never explores.
 
    It will always pick the action with the highest value from the Q-table. While these values will be updated, it
    never explores, so will likely quickly converge on a single action.
    """
 
    def __init__(self, k: int, start_value: float = 0.0) -> None:
        """
        Construct the agent.
 
        @param k The number of arms to select from. Should be an int greater than zero.
        @param start_value The starting reward to use for each arm. All arms assume the same value at the start.
        """
        super().__init__(k, start_value=start_value)
        # Track how many selections have been made to use in the update formula.
        self._n = 0
 
    def act(self) -> int:
        """
        Select an action to take from the available ones.
 
        Greedy always exploits, so this will always be one of the actions with the highest table value.
        @return An int representing the selected action. It will be on the interval [0, k).
        """
        return self.exploit()
 
    def update(self, action: int, reward: float) -> None:
        """
        Update the table values based on the last action.
 
        This uses an iterative version of a running average to update table values.
        @param action The index corresponding to the action that was taken.
        @param reward The resulting reward that was earned.
        """
        self._n += 1
        self.table[action] += (reward - self.table[action]) / self._n