× type: MemoryGym
name: SearingSpotlights-v0
frame_skip: 1
last_action_to_obs: False
last_reward_to_obs: False
obs_stacks: 1
grayscale: False
resize_vis_obs: [84, 84]
positional_encoding: False
reset_params:
        start-seed: 200001
        num-seeds: 1
        max_steps: 256
        initial_spawns: 4
        num_spawns: 30
        initial_spawn_interval: 30
        spawn_interval_threshold: 10
        spawn_interval_decay: 0.95
        spot_min_radius: 7.5
        spot_max_radius: 13.75
        spot_min_speed: 0.0025
        spot_max_speed: 0.0075
        spot_damage: 1.0
        visual_feedback: True
        black_background: False
        hide_chessboard: False
        light_dim_off_duration: 6
        light_threshold: 255
        num_coins: [1]
        coin_scale: 0.375
        coins_visible: False
        use_exit: True
        exit_visible: False
        exit_scale: 0.5
        agent_speed: 3.0
        agent_health: 5
        agent_scale: 0.25
        agent_visible: False
        sample_agent_position: True
        show_last_action: True
        show_last_positive_reward: True
        reward_inside_spotlight: 0.0
        reward_outside_spotlight: 0.0
        reward_death: 0.0
        reward_exit: 1.0
        reward_max_steps: 0.0
        reward_coin: 0.25
        seed: 200001
reward_normalization: 0

× load_model: False
model_path:
checkpoint_interval: 500
activation: relu
vis_encoder: cnn
vec_encoder: linear
num_vec_encoder_units: 128
hidden_layer: default
num_hidden_layers: 1
num_hidden_units: 384
transformer:
        num_blocks: 3
        embed_dim: 384
        num_heads: 4
        memory_length: 256
        positional_encoding: absolute
        layer_norm: pre
        init_weights: xavier
        gtrxl: False
        gtrxl_bias: 0.0
        gtrxl_swap: False
        share_heads: True
        max_episode_steps: 257
        add_positional_encoding_to_query: False
obs_decoder:
        attach_to: memory
        detach_gradient: False

× n_workers: 32
worker_steps: 512

× algorithm: PPO
resume_at: 0
gamma: 0.995
lamda: 0.95
updates: 30000
epochs: 3
refresh_buffer_epoch: -1
n_mini_batches: 8
advantage_normalization: no
value_coefficient: 0.5
max_grad_norm: 0.25
share_parameters: True
learning_rate_schedule:
        initial: 0.000275
        final: 1e-05
        power: 1.0
        max_decay_steps: 20000
beta_schedule:
        initial: 0.0001
        final: 1e-06
        power: 1.0
        max_decay_steps: 20000
clip_range_schedule:
        initial: 0.1
        final: 0.1
        power: 1.0
        max_decay_steps: 10000
obs_reconstruction_schedule:
        initial: 0.1
        final: 0.1
        power: 1.0
        max_decay_steps: 1000
ground_truth_estimator_schedule: {'initial': 0.0}