DOOM-RL TACTICAL MONITOR
Teaching an AI to Survive Doom — From Random Chaos to Combat Proficiency
[LEVEL_SELECT_SCREEN]
SCENARIO_1: BASIC
★☆☆☆☆DIFFICULTYEASY
TRAINED_EPISODES1,000
FINAL_MEAN_REWARD-11.1
ACTION_SPACE_SIZE5 ACTIONS
SCENARIO_2: DEFEND_CENTER
★★★☆☆DIFFICULTYMEDIUM
TRAINED_EPISODES2,000
FINAL_MEAN_REWARD+6.9
ACTION_SPACE_SIZE5 ACTIONS
SCENARIO_3: DEADLY_CORRIDOR
★★★★★DIFFICULTYNIGHTMARE
TRAINED_EPISODES3,000
FINAL_MEAN_REWARD-127.4
ACTION_SPACE_SIZE11 ACTIONS
[ANALYTIC_CHARTS]
[AGENT_PLAYBACK_COMPARISON]
RANDOM_AGENT (EPISODE 1)
TRAINED_AGENT (OPTIMAL)
[DQN_CONVOLUTIONAL_NEURAL_NETWORK]
INPUT
4x84x84
Stacked Frames
CONV_1
32 @ 8x8
Stride 4 + ReLU
CONV_2
64 @ 4x4
Stride 2 + ReLU
CONV_3
64 @ 3x3
Stride 1 + ReLU
DENSE_1
512 Units
Fully Connected
OUTPUT
5 Q-Values
Actions
[RESEARCH_FOUNDATION]
Mnih et al. (2013) — "Playing Atari with Deep Reinforcement Learning"
This project replicates the core convolutional neural network architecture presented by DeepMind in 2013, applying it to high-speed 3D spatial navigation tasks. We extend the baseline framework by implementing local configuration abstractions, shaped continuous distance metrics to bypass rewards sparseness, and custom action combinations mapping multiple buttons to unified inputs.
[SYSTEM_HARDWARE_SPECS]
| TRAINING_GPU | NVIDIA GeForce RTX 5080 (16GB VRAM, SM 12.0) |
| CUDA_VERSION | 12.8 |
| PYTORCH_VERSION | 2.10.0+cu128 |
| AVERAGE_TRAINING_FPS | ~550-600 steps/sec (rtx 5080 backpropagation) |
| TOTAL_EXPERIENCES_STORED | 110,000 steps across buffers |