aixi.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
A script for running AIXI-based agents in an environment, as configured by given options or the
given configuration file.

Usage: python aixi.py [-a | --agent <agent module name>]
                      [-d | --explore-decay <exploration decay value, between 0 and 1>]
                      [-e | --environment <environment module name>]
                      [-h | --agent-horizon <search horizon>]
                      [-l | --learning-period <cycle count>]
                      [-m | --mc-simulations <number of simulations to run each step>]
                      [-o | --option <extra option name>=<value>]
                      [-p | --profile]
                      [-r | --terminate-age <number of cycles before stopping the run>]
                      [-t | --ct-depth <maximum depth of predicting context tree>]
                      [-x | --exploration <exploration factor, greater than 0>]
                      [-v | --verbose]
                      [<environment configuration file name to load>]
"""

from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import os
import sys

import six.moves.configparser as configparser

try:
    import cProfile as profile
except:
    import profile
# end try

try:
    import cStringIO as StringIO
except:
    try:
        import StringIO
    except:
        import io as StringIO
    # end try
# end try

import datetime
import getopt
import inspect
import logging
import random
import sys
import time

# Insert the current directory into the system search path, so that this package can be
# imported when this script is run directly from a release archive.
PROJECT_ROOT = os.path.realpath(os.curdir)
sys.path.insert(0, PROJECT_ROOT)

from pyaixi import agent, agents, environment, environments, util

from pyaixi.agent import Agent
from pyaixi.agents import *
from pyaixi.environment import Environment
from pyaixi.environments import *

def interaction_loop(agent = None, environment = None, options = {}):
    """ The main agent/environment interaction loop.

        Each interaction cycle begins with the agent receiving an
        observation and reward from the environment.

        Subsequently, the agent selects an action and informs the environment.

        The interactions that took place are logged to the logger.
        When the cycle equals a power of two, a summary of the interactions is printed to
        the standard output.

        - `agent`: the agent object.
        - `environment`: the environment object.
        - `options`: the configuration options.

        (Called `mainLoop` in the C++ version.)
    """

    # Apply a random seed (Default: 0)
    random.seed(int(options.get("random-seed", 0)))

    # Verbose output (Default: False)
    verbose = bool(options.get("verbose", False))

    # Determine exploration options. (Default: don't explore, don't decay.)
    explore_rate = float(options.get("exploration", 0.0))
    explore = (explore_rate > 0)
    explore_decay = float(options.get("explore-decay", 1.0))
    assert 0.0 <= explore_rate
    assert 0.0 <= explore_decay and explore_decay <= 1.0

    # Determine termination age. (Default: don't terminate)
    terminate_age = int(options.get("terminate-age", 0))
    terminate_check = (terminate_age > 0)
    assert 0 <= terminate_age

    # Determine the cycle after which the agent stops learning (if ever.)
    learning_period = int(options.get("learning-period", 0))
    assert 0 <= learning_period

    # Agent/environment interaction loop.
    cycle = 1
    while not environment.is_finished:
        # Check for agent termination.
        if terminate_check and agent.age > terminate_age:
            break
        # end if

        # Save the current time to compute how long this cycle took.
        cycle_start = datetime.datetime.now()

        # Get a percept from the environment.
        observation = environment.observation
        reward = environment.reward

        # If we're outside the learning period, stop exploring.
        if learning_period > 0 and cycle > learning_period:
            explore = False
        # end if

        # Update the agent's environment model with the new percept.
        agent.model_update_percept(observation, reward)

        # Determine best exploitive action, or explore.
        explored = False
        if explore and (random.random() < explore_rate):
            # Yes, we're still exploring.
            # Generate a random action to explore.
            explored = True
            if verbose:
                # Tell the user the agent is exploring at random.
                print("Agent is trying an action at random...")
            # end if
            action = agent.generate_random_action()
        else:
            # No, we're not still exploring.
            # Exploit our past learning to work out the best action.
            if verbose:
                # Tell the user we're not exploring, we're trying to choose the best action.
                print("Agent is trying to choose the best action, which may take some time...")
            # end if
            action = agent.search()
        # end def

        # Send the action to the environment.
        environment.perform_action(action)

        # Update the agent's environment model with the chosen action.
        agent.model_update_action(action)

        # Calculate how long this cycle took.
        time_taken = datetime.datetime.now() - cycle_start

        # Log this cycle.
        message = "%d, %s, %s, %s, %s, %f, %d, %f, %s, %d" % \
                  (cycle, str(observation), str(reward),
                   str(action), str(explored), explore_rate,
                   agent.total_reward, agent.average_reward(),
                   str(time_taken), agent.model_size())
        print(message)

        # Print to standard output when cycle == 2^n or on verbose option.
        if verbose or (cycle & (cycle - 1)) == 0:
            message = "cycle: %s" % str(cycle) + os.linesep + \
                      "average reward: %f" % agent.average_reward()
            if explore:
                message += os.linesep + "explore rate: %f" % float(explore_rate) + os.linesep
            # end if

            print(message)
        # end def

        # Print environment state if verbose option is true.
        if verbose:
              print(environment.print())
        # end if

        # Update exploration rate.
        if explore:
            explore_rate *= explore_decay
        # end def

        # Update the cycle count.
        cycle += 1
    # end while

    # Print summary to standard output.
    message = "SUMMARY:" + os.linesep + \
              "agent age: %d" % agent.age + os.linesep + \
              "average reward: %f" % agent.average_reward()

    print(message)
# end def

def main(argv):
    """ Entry point of the program. Sets up logging, default configuration values,
        environment and agent before starting the agent/environment interaction cycle
        by calling `interaction_loop`.

        If invalid arguments or options are given, it prints usage help information
        to the standard output and exits.
    """

    # Define some default configuration values.
    default_options = {}
    default_options["agent"]           = "mc_aixi_ctw"
    default_options["agent-horizon"]   = 5
    default_options["ct-depth"]        = 30
    default_options["environment"]     = "coin_flip"
    default_options["exploration"]     = 0.0    # Do not explore.
    default_options["explore-decay"]   = 1.0    # Exploration rate does not decay.
    default_options["learning-period"] = 0      # Learn forever.
    default_options["mc-simulations"]  = 300
    default_options["profile"]         = False  # Whether to profile code.
    default_options["terminate-age"]   = 0      # Never die.
    default_options["verbose"]         = False

    command_line_options = {}

    # Process the command line options and arguments.
    try:
        opts, args = getopt.gnu_getopt(
                                       argv,
                                       'd:e:h:l:m:o:pr:t:vx:',
                                       ['explore-decay=', 'environment=', 'agent-horizon=',
                                        'learning-period=', 'mc-simulations=', 'option', 'profile',
                                        'terminate-age=', 'ct-depth=', 'verbose', 'exploration=',]
                                      )

        for opt, arg in opts:
            if opt == '--help':
                usage()
            # end if
            if opt in ('-d', '--explore-decay'):
                command_line_options["explore-decay"] = float(arg)
                continue
            # end if
            if opt in ('-e', '--environment'):
                command_line_options["environment"] = str(arg)
                continue
            # end if
            if opt in ('-h', '--agent-horizon'):
                command_line_options["agent-horizon"] = int(arg)
                continue
            # end if
            if opt in ('-l', '--learning-period'):
                command_line_options["learning-period"] = int(arg)
                continue
            # end if
            if opt in ('-m', '--mc-simulations'):
                command_line_options["mc-simulations"] = int(arg)
                continue
            # end if
            if opt in ('-o', '--option'):
                # Split the associated argument into a key and value pair, splitting on the '=' symbol.
                parts = arg.split("=")

                # Do we have enough parts to make a key=value pair?
                if len(parts) > 1:
                    key = parts[0].strip()
                    value = '='.join(parts[1:])
                    command_line_options[key] = value
                else:
                    # No. Show the usage, after printing an explantory message.
                    print("Extra option '-o %s' is invalid. " % str(arg) + \
                          "This needs to be in '-o key=value' format." % str(arg))
                    usage()
                # end if
                continue
            # end if
            if opt in ('-p', '--profile'):
                command_line_options["profile"] = True
                continue
            # end if
            if opt in ('-r', '--terminate-age'):
                command_line_options["terminate-age"] = int(arg)
                continue
            # end if
            if opt in ('-t', '--ct-depth'):
                command_line_options["ct-depth"] = int(arg)
                continue
            # end if
            if opt in ('-v', '--verbose'):
                command_line_options["verbose"] = True
                continue
            # end if
            if opt in ('-x', '--exploration'):
                command_line_options["exploration"] = float(arg)
                continue
            # end if
        # end for
    except getopt.GetoptError as e:
        # We got an incorrect option. Show the usage and exit.
        usage()
    # end try

    # Do we have any arguments left over?
    if len(args) > 0:
        # Yes. The first should be the name of a configuration file.
        filename = args[0]

        # Is this a valid filename?
        if not os.path.exists(filename):
            print("Expected argument '%s' to be a configuration filename." % str(filename))
            usage()
        # end if

        # If we're here, we've got a valid filename.
        # Try reading it in as configuration file.
        config_contents = open(filename, 'r').read()

        # Does the configuration contents contain an 'environment' section?
        if config_contents.find("[environment]") == -1:
            # No. Add one to the beginning.
            config_contents = "[environment]" + os.linesep + config_contents
        # end if

        # Convert the contents into an in-memory file-like object, for parsing.
        config_stringio = StringIO.StringIO(config_contents)

        # Parse the given options, giving the default options as defaults to the parser.
        config = configparser.RawConfigParser(default_options)
        config.readfp(config_stringio)

        # Get the configuration options read in as a dictionary.
        # (This should exist in a section called 'environment'.)
        options = dict(config.items('environment'))
    else:
        # No. So set the options to be the default options.
        options = default_options
    # end if

    # Let the command line options override the options read from the configuration file--or
    # the default values--whichever way we got to this point.
    options.update(command_line_options)

    # Print the options we've received, if we've been requested to be verbose.
    verbose = bool(options.get("verbose", False))
    if verbose:
        for option_name, option_value in list(options.items()):
            print("OPTION: '%s' = '%s'" % (str(option_name), str(option_value)))
        # end for
    # end if

    # Print an initial message header.
    message = "cycle, observation, reward, action, explored, " + \
              "explore_rate, total reward, average reward, time, model size"
    print(message)

    # Try to import an agent module with the given name.
    agent_name = options["agent"]

    # Ensure the name of the package we're trying to import has a prefix of 'pyaixi.agents',
    # if it doesn't have one specified already.
    if agent_name.count('.') == 0:
        agent_package_name = "pyaixi.agents." + agent_name
    else:
        agent_package_name = agent_name
    # end if

    try:
        agent_module = __import__(agent_package_name, globals(), locals(), [agent_name], 0)
    except Exception as e:
        # Exit with an error.
        sys.stderr.write("ERROR: loading agent module '%s' caused error '%s'. Exiting." % \
                         (str(agent_name), str(e)) + os.linesep)
        sys.exit(1)
    # end try

    # Find a subclass of the Agent class in the given module.
    agent_class = None
    agent_classname = ""
    for name in dir(agent_module):
        obj = getattr(agent_module, name)
        if inspect.isclass(obj) and 'Agent' in [cls.__name__ for cls in obj.__bases__]:
            agent_class = obj
            agent_classname = name
            break
        # end if
    # end for

    # Did we find a subclass of Agent?
    if agent_class is None:
        # No. Exit with an error.
        sys.stderr.write("ERROR: agent module '%s' does not contain " % str(agent_name) + \
                         "a valid AIXI agent subclass. (Got '%s' instead.) Exiting." % \
                         str(agent_classname) + os.linesep)
        sys.exit(1)
    # end if

    # Try to import an environment module with the given name.
    environment_name = options["environment"]

    # Ensure the name of the package we're trying to import has a prefix of 'pyaixi.environments',
    # if it doesn't have one specified already.
    if environment_name.count('.') == 0:
        environment_package_name = "pyaixi.environments." + environment_name
    else:
        environment_package_name = environment_name
    # end if

    try:
        environment_module = __import__(environment_package_name, globals(), locals(),
                                        [environment_name], 0)
    except Exception as e:
        # Exit with an error.
        sys.stderr.write("ERROR: loading environment module '%s' caused error '%s'. Exiting." % \
                         (str(environment_name), str(e)) + os.linesep)
        sys.exit(1)
    # end try

    # Find a subclass of the Environment class in the given module.
    environment_class = None
    environment_classname = ""
    for name, obj in inspect.getmembers(environment_module):
        if hasattr(obj, "__bases__") and 'Environment' in [cls.__name__ for cls in obj.__bases__]:
            environment_class = obj
            environment_classname = name
            break
        # end if
    # end for

    # Did we find a subclass of Environment?
    if environment_class is None:
        # No. Exit with an error.
        sys.stderr.write("ERROR: environment module '%s' does not contain " % str(environment_name) + \
                         "a valid AIXI environment subclass. (Got '%s' instead.) Exiting." % \
                          str(environment_classname) + os.linesep)
        sys.exit(1)
    # end if

    # Create an instance of the environment, using the discovered options.
    environment = environment_class(options = options)

    # Copy environment-dependent configuration options to the options.
    options["action-bits"] = environment.action_bits()
    options["observation-bits"] = environment.observation_bits()
    options["percept-bits"] = environment.percept_bits()
    options["reward-bits"] = environment.reward_bits()
    options["max-action"] = environment.maximum_action()
    options["max-observation"] = environment.maximum_observation()
    options["max-reward"] = environment.maximum_reward()

    # Set up the agent, using the created environment, and the updated options.
    agent = agent_class(environment = environment, options = options)

    # Run the main agent/environment interaction loop, profiling if requested to do so.
    if bool(options.get("profile", False)):
        profile.runctx('interaction_loop(agent = agent, environment = environment, options = options)',
                       globals(), locals())
    else:
        interaction_loop(agent = agent, environment = environment, options = options)
    # end def
# end def

def usage():
    """ Prints usage information.
    """

    message = "Usage: python aixi.py [-a | --agent <agent module name>" + os.linesep + \
              "                      [-d | --explore-decay <exploration decay value, between 0 and 1>]" + os.linesep + \
              "                      [-e | --environment <environment module name>]" + os.linesep + \
              "                      [-h | --agent-horizon <search horizon>]" + os.linesep + \
              "                      [-l | --learning-period <cycle count>]" + os.linesep + \
              "                      [-m | --mc-simulations <number of simulations to run each step>]" + os.linesep + \
              "                      [-o | --option <extra option name>=<value>]" + os.linesep + \
              "                      [-p | --profile]" + os.linesep + \
              "                      [-r | --terminate-age <number of cycles before stopping the run>]" + os.linesep + \
              "                      [-t | --ct-depth <maximum depth of predicting context tree>]" + os.linesep + \
              "                      [-x | --exploration <exploration factor, greater than 0>]" + os.linesep + \
              "                      [-v | --verbose]" + os.linesep + \
              "                      [<configuration file name to load>]" + os.linesep + os.linesep

    sys.stderr.write(message)
    sys.exit(2)
# end def


# Start the main function if this file has been executed, and not just imported.
if __name__ == "__main__":
    main(sys.argv[1:])
# end def