- TORCS is a continuous control problem. DQN works only for discrete actions, and so it cannot be used in TORCS.
- The initialization is another initialization strategy; you can also use a random uniform initialization with the min and max values of the range specified; another approach is to sample from a Gaussian with a zero mean and a specified sigma value. The interested reader must try these different initializers and compare the agent's performance.
- The abs() function is used in the reward function, as we penalize lateral drift from the center equally on either side (left or right). The first term is the longitudinal speed, and so no abs() function is required.
- The Gaussian noise added to the actions for exploration can be tapered down with episode count, and this can result in smoother driving. Surely, there are many other tricks you can do!
- DDPG is off-policy...
Germany
Slovakia
Canada
Brazil
Singapore
Hungary
Philippines
Mexico
Thailand
Ukraine
Luxembourg
Estonia
Lithuania
Norway
Chile
United States
Great Britain
India
Spain
South Korea
Ecuador
Colombia
Taiwan
Switzerland
Indonesia
Cyprus
Denmark
Finland
Poland
Malta
Czechia
New Zealand
Austria
Turkey
France
Sweden
Italy
Egypt
Belgium
Portugal
Slovenia
Ireland
Romania
Greece
Argentina
Malaysia
South Africa
Netherlands
Bulgaria
Latvia
Australia
Japan
Russia