yawning_titan.envs.generic.helpers.eval_printout

Util to print out agent evaluation metrics.

The metrics printed out are:

• Total episodes elapsed

• Absolute wins for red and blue

• Percentage win rate for red and blue

• Average episode length

• Actions taken by blue each game/Average actions taken by blue over n games

Classes

EvalPrintout

Class to represnt an Eval Printer.

class yawning_titan.envs.generic.helpers.eval_printout.EvalPrintout(avg_every)

Class to represnt an Eval Printer.

Initialise printout object.

Parameters:

avg_every – Number of timesteps to average stats over

__init__(avg_every)

Initialise printout object.

Parameters:

avg_every – Number of timesteps to average stats over

print_stats(game_stats_list, total_games)

Print out the (averaged) stats from the last avg_every number of games to the console.

Parameters:

• game_stats_list – List of dictionaries containing the last avg_every number of game stats

• total_games – Total games played since starting

calculate_metrics(game_stats_list)

Calculate the metrics to be printed.

Parameters:

game_stats_list – List of dictionaries containing the last avg_every number of game stats

Returns:

Number of games blue won in the last avg_every number of games red_wins: Number of games red won in the last avg_every number of games percentage_blue: Percentage of games blue won in the last avg_every number of games percentage_red: Percentage of games red won in the last avg_every number of games avg_duration: Average number of timesteps per episode over the last avg_every number of games sorted_actions: Dictionary of actions taken by blue, averaged over the last avg_every number of games

and ordered by frequency of each action from highest to lowest. Dictionary values are tuples: (average frequency of action, action usage percentage)

Return type:

blue_wins