add md

app.py

@@ -52,7 +52,19 @@ def get_frame_and_attribution(slider_value):
     
 with gr.Blocks() as demo:
     gr.Markdown("# Introspection in Deep Reinforcement Learning")
-
+    gr.Markdown(r"""
+    \#\# How this space works:
+    This space was created for trying to apply [Integrated Gradients](https://captum.ai/docs/extension/integrated_gradients\#:~:text=Integrated%20gradients%20is%20a%20simple,and%20feature%20or%20rule%20extraction.) \
+                into Deep Reinforcement Learning Scenarions. It uses PyTorch's captum library for interpretability, and Gymnasium for the emulator of the continuous lunar lander.
+    \#\#\# Training algorithm: [DDPG](https://arxiv.org/abs/1509.02971)
+                This agent was trained with Deep Deterministic Policy Gradients, and outputs an average reward of 260.8 per episode (successful)
+    \#\#\# Using this space:
+                - First, select the environment (futurely there will be more environments available)
+                - Then, select if you want the baseline (see IG paper for more detail) to be \
+                a torch `tensor` of zeroes, or a running average of the initial frames of a few episodes (selected on the right) \
+                - Click attribute and wait a few seconds (usually 20-25s) for the attributions to be computed with the trained agent over 10 episodes
+                - Finally, use the slider to get a key frame that tells the attributions of the agent. They're under a Softmax to fit the component's requirements for a probability distribution. 
+    """)
     with gr.Tab(label="Attribute"):
         env_spec = gr.Dropdown(choices=["LunarLander-v2"],type="value",multiselect=False, label="Environment Specification (e.g.: LunarLander-v2)")
         env = gr.Interface(title="Create the Environment", allow_flagging="never", inputs=env_spec, fn=create_training_loop, outputs=gr.JSON())
@@ -64,5 +76,20 @@ with gr.Blocks() as demo:
         slider = gr.Slider(label="Key Frame", minimum=0, maximum=1000, step=1, value=0)
 
         gr.Interface(fn=get_frame_and_attribution, inputs=slider, live=True, outputs=[gr.Image(label="Timestep"),gr.Label(label="Attributions")])
+    gr.Markdown(r"""\#\# Local Usage and Packages \
+    `pip install torch gymnasium 'gymnasium[box2d]'` \
+    You might need to install Box2D Separately, which requires a swig package to compile code from Python into C/C++, which is the language that Box2d was built in: \
+    `brew install swig` \
+    `pip install box2d \n \#\# Average Score: 164.38 (significant improvement from discrete action spaces) \
+    For each step, the reward: \
+    - is increased/decreased the closer/further the lander is to the landing pad. \
+    - is increased/decreased the slower/faster the lander is moving.\
+    - is decreased the more the lander is tilted (angle not horizontal). \
+    - is increased by 10 points for each leg that is in contact with the ground. \
+    - is decreased by 0.03 points each frame a side engine is firing.\
+    - is decreased by 0.3 points each frame the main engine is firing. \
+    The episode receives an additional reward of -100 or +100 points for crashing or landing safely respectively. An episode is considered a solution if it scores at least 200 points.\*\* \
+    \#\# `train()` and `load_trained()` \
+    `load_trained()` function loads a pre-trained model that ran through 1000 episodes of training, while `train()` does training from scratch. You can edit which one of the functions is running from the bottom of the main.py file. If you set render_mode=False, the program will train a lot faster.)\n demo.launch()""")
 
 demo.launch()