csle_agents.agents.ppo_clean package

Submodules

csle_agents.agents.ppo_clean.ppo_clean_agent module

MIT License

Copyright (c) 2019 CleanRL developers https://github.com/vwxyzjn/cleanrl

class csle_agents.agents.ppo_clean.ppo_clean_agent.PPOCleanAgent(simulation_env_config: csle_common.dao.simulation_config.simulation_env_config.SimulationEnvConfig, emulation_env_config: Union[None, csle_common.dao.emulation_config.emulation_env_config.EmulationEnvConfig], experiment_config: csle_common.dao.training.experiment_config.ExperimentConfig, training_job: Optional[csle_common.dao.jobs.training_job_config.TrainingJobConfig] = None, save_to_metastore: bool = True)[source]

Bases: csle_agents.agents.base.base_agent.BaseAgent

A PPO agent using the implementation from CleanRL

generalized_advantage_estimation(model: csle_common.models.ppo_network.PPONetwork, next_obs: torch.Tensor, rewards: torch.Tensor, device: torch.device, next_done: torch.Tensor, dones: torch.Tensor, values: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor][source]

Computes the generalized advantage estimation (i.e., exponentially weighted average of n-step returns) See (HIGH-DIMENSIONAL CONTINUOUS CONTROL USING GENERALIZED ADVANTAGE ESTIMATION, 2016, ICLR)

Parameters

  • device – the device acted upon

  • values – tensor of values

  • model – the neural network model

  • rewards – tensor of available rewards

  • next_obs – the next observation

  • dones – tensor of done events

  • next_done – logical or operation between terminations and truncations

Returns

returns, advantages

hparam_names() → List[str][source]
Returns

a list with the hyperparameter names

make_env() → Callable[[], gymnasium.wrappers.record_episode_statistics.RecordEpisodeStatistics][source]

Helper function for creating the environment to use for training

Returns

a function that creates the environment

run_ppo(exp_result: csle_common.dao.training.experiment_result.ExperimentResult, seed: int) → Tuple[csle_common.dao.training.experiment_result.ExperimentResult, csle_common.dao.simulation_config.base_env.BaseEnv, csle_common.models.ppo_network.PPONetwork][source]

Runs PPO with a given seed

Parameters

  • exp_result – the object to save the experiment results

  • seed – the random seed

Returns

the updated experiment results, the environment, and the trained model

train() → csle_common.dao.training.experiment_execution.ExperimentExecution[source]

Runs the training process

Returns

the results

update_trajectory_buffers(global_step: int, envs: gymnasium.vector.sync_vector_env.SyncVectorEnv, obs: torch.Tensor, dones: torch.Tensor, actions: torch.Tensor, rewards: torch.Tensor, device: torch.device, logprobs: torch.Tensor, values: torch.Tensor, model: csle_common.models.ppo_network.PPONetwork, next_obs: torch.Tensor, next_done: torch.Tensor, horizons: List[int]) → Tuple[torch.Tensor, torch.Tensor, int, List[int]][source]

Updates the buffers of trajectories collected from the environment

Parameters

  • step (global) – the global step in the iteration

  • envs – list of environments

  • obs – torch tensor of observations

  • dones – tensor of done events

  • actions – tensor of available actions

  • rewards – tensor of available rewards

  • device – the device acted upon

  • logprobs – logarithmic probabilities

  • horizons – list of time horizons

  • values – tensor of values

  • model – the neural network model

  • next_obs – the next observation

  • next_done – logical or operation between terminations and truncations

Returns

next_obs, next_done, global_step

Module contents