Module TeachMyAgent.teachers.utils.alpha_functions

Expand source code 
# Taken from https://github.com/psclklnk/spdl
# Copy of the license at TeachMyAgent/teachers/LICENSES/SPDL
from abc import ABC, abstractmethod
import torch


class AlphaFunction(ABC):

    @abstractmethod
    def __call__(self, iteration, average_reward, kl_divergence):
        pass


class PercentageAlphaFunction(AlphaFunction):
    def __init__(self, offset, percentage):
        '''
            Calculate an automatically adjusted alpha parameter to maintain constant proportion.

            :param percentage: proportion to maintain
            :param offset: How many times alpha should be set to 0
        '''
        self.offset = offset
        self.percentage = percentage

    def __call__(self, iteration, average_reward, kl_divergence):
        if iteration < self.offset:
            alpha = 0.
        else:
            kl_divergence = torch.clamp(kl_divergence, min=1e-10)
            average_reward = 0. if average_reward < 0. else average_reward
            alpha = torch.clamp(self.percentage * average_reward / kl_divergence, max=1e5)
        return alpha

Classes

class AlphaFunction

Helper class that provides a standard way to create an ABC using inheritance.

Expand source code 
class AlphaFunction(ABC):

    @abstractmethod
    def __call__(self, iteration, average_reward, kl_divergence):
        pass

Ancestors

  • abc.ABC

Subclasses

class PercentageAlphaFunction (offset, percentage)

Helper class that provides a standard way to create an ABC using inheritance.

Calculate an automatically adjusted alpha parameter to maintain constant proportion.

:param percentage: proportion to maintain :param offset: How many times alpha should be set to 0

Expand source code 
class PercentageAlphaFunction(AlphaFunction):
    def __init__(self, offset, percentage):
        '''
            Calculate an automatically adjusted alpha parameter to maintain constant proportion.

            :param percentage: proportion to maintain
            :param offset: How many times alpha should be set to 0
        '''
        self.offset = offset
        self.percentage = percentage

    def __call__(self, iteration, average_reward, kl_divergence):
        if iteration < self.offset:
            alpha = 0.
        else:
            kl_divergence = torch.clamp(kl_divergence, min=1e-10)
            average_reward = 0. if average_reward < 0. else average_reward
            alpha = torch.clamp(self.percentage * average_reward / kl_divergence, max=1e5)
        return alpha

Ancestors