UCAS course[Reinforcement Learning] group assignment, as well as the materious for COG competition 2022.
离散的DQN
CogEnvDecoder == 0.1.29
discrete_v2.py 中 修改了小车朝向与目标之间的夹角；奖励函数是 欧氏距离+碰撞次数+碰撞时间；训练过程中每找到一个目标就reset。
test_env.py 是直接evaluate上面discrete_v2.py中的最好的模型
continue.py 三个动作 dx dy dtheta; 奖励函数：goal_reward/ dis_reward/ collision_reward； 训练方式：遇到障碍物就reset； 每局训练
存在问题：有时会卡在某一个地方。
continue_v3.py 设置卡顿标志，加入临时goal：卡顿后，有30步前往临时目标； 每步训练
test_env_continue.py 环境测试 v2版本，要提前在官网的 tag 里下载好
continue_v4.py: update the flag of stuck.
Toolbox.py 中提供了更新的计算距离的方式：高斯混合模型模拟障碍物 曲线积分计算实际距离
We accidently find that the rotation is not an absolute bad state actually. By set the reward function to "100 * base_reward + 10 * rotate_loss + 1000 * collision_loss" and train the agent with 100000 steps, we get a great agent which could succeed in finding the object about 4-5 times averagely in 10 repeated evaluations. The rotated agent perform well probably because the agent stucks with a lower probability compared with those who do not rotate. Even to the extent that the rotation could prevent the agent from stuck. We still find that the performance went bad with more training steps. Moreover, we get some tips about the influence on the performance by adjust the reward function. With a higher weights for rotate_loss, the agent stops rotating at an earlier traing step.