def test_agent_training() -> None:
dqn_learning_run = learner.epsilon_greedy_search(
cyberbattle_gym_env=cyberbattlechain,
environment_properties=ep,
learner=dqla.DeepQLearnerPolicy(
ep=ep,
gamma=0.015,
replay_memory_size=10000,
target_update=10,
batch_size=512,
learning_rate=0.01), # torch default is 1e-2
episode_count=training_episode_count,
iteration_count=iteration_count,
epsilon=0.90,
render=False,
# epsilon_multdecay=0.75, # 0.999,
epsilon_exponential_decay=5000, # 10000
epsilon_minimum=0.10,
verbosity=Verbosity.Quiet,
title="DQL")
assert dqn_learning_run
random_run = learner.epsilon_greedy_search(
cyberbattlechain,
ep,
learner=learner.RandomPolicy(),
episode_count=training_episode_count,
iteration_count=iteration_count,
epsilon=1.0, # purely random
render=False,
verbosity=Verbosity.Quiet,
title="Random search")
assert random_run
cyberbattlechain_10,
ep,
learner=best_dqn_learning_run_10['learner'],
episode_count=eval_episode_count,
iteration_count=iteration_count,
epsilon=0.0, # 0.35,
render=False,
render_last_episode_rewards_to='images/chain10',
title="Exploiting DQL",
verbosity=Verbosity.Quiet)
# %%
random_run = learner.epsilon_greedy_search(
cyberbattlechain_10,
ep,
learner=learner.RandomPolicy(),
episode_count=eval_episode_count,
iteration_count=iteration_count,
epsilon=1.0, # purely random
render=False,
verbosity=Verbosity.Quiet,
title="Random search")
# %%
# Plot averaged cumulative rewards for DQL vs Random vs DQL-Exploit
themodel = dqla.CyberBattleStateActionModel(ep)
p.plot_averaged_cummulative_rewards(
all_runs=[best_dqn_learning_run_10, random_run, dql_exploit_run],
title=
f'Benchmark -- max_nodes={ep.maximum_node_count}, episodes={eval_episode_count},\n'
f'State: {[f.name() for f in themodel.state_space.feature_selection]} '