GitLab

- Also added ManualWait class.

- Added RestrictedEpsGreedyPolicy and RestrictedGreedyPolicy and use them as policy and test_policy in DQNLearner. Now, the agent never chooses the action corresponding to -inf Q-value if there is at least one action with finite Q-value (if not, it chooses any action randomly, which is necessary for compatibility with keras-rl --
 see the comments in select_action).

- Now, generate_scenario in SimpleIntersectionEnv generates veh_ahead_scenario even when randomize_special_scenario = 1.

- In EpisodicEnvBase, the terminal reward is by default determined by the minimum one;

- Small change of initiation_condition of EpisodicEnvBase (simplified);

The high-level policy was trained without changelane maneuver but with immediatestop maneuver. Two problems remain: 1) the agent chooses changelane maneuver too frequently; 2) before the stop region, immediatestop maneuver works but was not chosen property after 2.5m the high-level policy training...

Each low level policy was retrained with better LTL conditions and rewards, some parts of which are also designed to encourage exploration (to prevent the vehicle from being stopped all the time).