evorl.algorithms.contrib.td3_v3

Module Contents

Classes

TD3Agent	The Agnet for TD3.
TD3NetworkParams	Contains training state for the learner.
TD3TrainMetric
TD3V3Workflow	The similar impl of TD3 in SB3 and CleanRL.

Functions

make_mlp_td3_agent

Data

MISSING_LOSS

API

evorl.algorithms.contrib.td3_v3.MISSING_LOSS

None

class evorl.algorithms.contrib.td3_v3.TD3Agent

Bases: evorl.agent.Agent

The Agnet for TD3.

actor_network: flax.linen.Module

None

clip_policy_noise: float

0.5

compute_actions(agent_state: evorl.agent.AgentState, sample_batch: evorl.sample_batch.SampleBatch, key: chex.PRNGKey) → tuple[evorl.types.Action, evorl.types.PolicyExtraInfo]

critic_network: flax.linen.Module

None

critics_in_actor_loss: str

‘first’

discount: float

0.99

evaluate_actions(agent_state: evorl.agent.AgentState, sample_batch: evorl.sample_batch.SampleBatch, key: chex.PRNGKey) → tuple[evorl.types.Action, evorl.types.PolicyExtraInfo]

exploration_epsilon: float

0.5

init(obs_space: evorl.envs.Space, action_space: evorl.envs.Space, key: chex.PRNGKey) → evorl.agent.AgentState

property normalize_obs

obs_preprocessor: Any

‘pytree_field(…)’

policy_noise: float

0.2

class evorl.algorithms.contrib.td3_v3.TD3NetworkParams

Bases: evorl.types.PyTreeData

Contains training state for the learner.

actor_params: evorl.types.Params

None

critic1_params: evorl.types.Params

None

critic2_params: evorl.types.Params

None

target_actor_params: evorl.types.Params

None

target_critic1_params: evorl.types.Params

None

target_critic2_params: evorl.types.Params

None

class evorl.algorithms.contrib.td3_v3.TD3TrainMetric

Bases: evorl.metrics.MetricBase

actor_loss: chex.Array

None

critic1_loss: chex.Array

None

critic2_loss: chex.Array

None

q1: chex.Array

None

q2: chex.Array

None

class evorl.algorithms.contrib.td3_v3.TD3V3Workflow(env: evorl.envs.Env, agent: evorl.agent.Agent, optimizer: optax.GradientTransformation, evaluator: evorl.evaluators.Evaluator, replay_buffer: evorl.replay_buffers.AbstractReplayBuffer, config: omegaconf.DictConfig)

Bases: evorl.algorithms.offpolicy_utils.OffPolicyWorkflowTemplate

The similar impl of TD3 in SB3 and CleanRL.

learn(state: evorl.types.State) → evorl.types.State

classmethod name()

step(state: evorl.types.State) → tuple[evorl.metrics.MetricBase, evorl.types.State]

evorl.algorithms.contrib.td3_v3.make_mlp_td3_agent(action_space: evorl.envs.Space, norm_layer_type: str = 'none', critic_hidden_layer_sizes: tuple[int] = (256, 256), actor_hidden_layer_sizes: tuple[int] = (256, 256), discount: float = 0.99, exploration_epsilon: float = 0.5, policy_noise: float = 0.2, clip_policy_noise: float = 0.5, critics_in_actor_loss: str = 'first', normalize_obs: bool = False)