minor fix in readme

README.md

@@ -15,7 +15,7 @@ In particular, the network here is much smaller and can be easily trained withou
 The agent learns to drive the car from pixels in a few hours and doesn't need any hand-crafted features.
 There are some minor environment specific tweaks for the car-racing but the base-agent doesn't know about car-racing.
 
-#### pre-trained agent
+## pre-trained agent
 The checkpoint provided in the repo used the default parameters
 specified in the runner/agent and 150000~ playing steps for learning.
 
@@ -26,15 +26,15 @@ https://youtu.be/CVZQOAlQib0
 The agent sometimes cuts corners but other than that it can drive flawlessly for minutes.
 There are some occasional mistakes though.
 
-#### Running instuctions
-Just clone the repo and use car_runner.py
+## Running instuctions
+Just clone the repo and use car_runner.py main.
 The settings are specified in the beginning of the runner.
 
 You can either train from scratch or load the existing checkpoint
 from this repo and see the agent driving somewhat properly right away.
 Or you can train the provided checkpoint more.
 
-#### Dependencies
+## Dependencies
 - Python 3.5 (will not work with python 2)
 - OpenAI Gym (the car-racing environment)
 - Tensorflow 1.0.0