How nicolassalat/dspro2 works
This repository is a Mario Kart Wii image-based reinforcement-learning agent. Its main workflow captures gameplay frames, selects driving actions with a distributional Q-network, and trains from replayed experience; a separate greedy evaluation loop reports rewards and finish rate. The graph also preserves the emulator-side state/action interface and reward shaping. Training and evaluation source excerpts show their dependencies, but the unsampled Dolphin environment implementation limits claims about its internal runtime wiring.
Main entry points: btr.py BTR Network · evaluate.py Evaluation Loop · training_process.py Training Process · neural_agent.py Neural Agent
Components 18
Emulator Environment
- Dolphin Environment
scripts/env/dolphin_env.py - Game Memory
scripts/env/game_memory.py - Action Space
scripts/env/actions.py - Frame Capture
scripts/env/dolphin_capture.py - Reward Shaping
scripts/env/reward_function.py
Agent Training
- Training Process
scripts/train/training_process.py - Neural Agent
scripts/train/neural_agent.py - Replay Buffer
scripts/train/neural_agent.py
Q Network
- BTR Network
btr/networks/btr.py - IMPALA Encoder
btr/encoders/impala_encoder.py - Encoder Base
btr/encoders/base_encoder.py - IQN Head
btr/networks/iqn.py - Value Streams
btr/networks/value_stream.py - Noisy Layers
btr/layers/noisy_linear.py
Evaluation
- Evaluation Launcher
scripts/eval/start_eval.py - Evaluation Loop
scripts/eval/evaluate.py
Other components
- Operator
- Dolphin Emulator
Connections 21
- Operator → Training Process: starts training
README.md - Training Process → Frame Capture: uses frames
scripts/train/training_process.py - Training Process → Reward Shaping: computes reward
scripts/train/training_process.py - Training Process → Neural Agent: trains agent
scripts/train/training_process.py - Neural Agent → Replay Buffer: stores transitions
scripts/train/neural_agent.py - Neural Agent → BTR Network: runs policy
scripts/train/neural_agent.py - BTR Network → IMPALA Encoder: encodes frames
btr/networks/btr.py - IMPALA Encoder → Encoder Base: extends
btr/encoders/impala_encoder.py - BTR Network → IQN Head: embeds quantiles
btr/networks/btr.py - BTR Network → Value Streams: combines streams
btr/networks/btr.py - Value Streams → Noisy Layers: uses layers
btr/networks/value_stream.py - BTR Network → Noisy Layers: resets noise
btr/networks/btr.py - Operator → Evaluation Launcher: starts evaluation
README.md - Evaluation Launcher → Evaluation Loop: launches loop
README.md - Evaluation Loop → Frame Capture: captures frames
scripts/eval/evaluate.py - Evaluation Loop → Reward Shaping: scores progress
scripts/eval/evaluate.py - Evaluation Loop → Neural Agent: selects actions
scripts/eval/evaluate.py - Evaluation Loop → Dolphin Emulator: accepts emulator links
scripts/eval/evaluate.py - Dolphin Emulator → Dolphin Environment: hosts environment
README.md - Dolphin Environment → Game Memory: reads state
scripts/env/dolphin_env.py - Dolphin Environment → Action Space: maps controls
scripts/env/dolphin_env.py