【回答问题】ChatGPT上线了!比较流行的强化学习算法

在这里插入图片描述
强化学习是人工智能和机器学习领域的一个重要分支。它主要研究的是如何让计算机在有目的的学习过程中自动找到最优的行动策略。

强化学习的基本过程是:环境与智能体之间不断进行交互,智能体根据环境的反馈来不断学习,并逐渐找到最优的行动策略。

在强化学习中,智能体的目标是获得最大的长期奖励,通过不断地尝试不同的行动来实现这一目标。每一次尝试都会产生一定的奖励或惩罚,而智能体要学会根据这些奖励来更新自己的行动策略,从而达到获得最大奖励的目标。

在强化学习的过程中,智能体可以使用不同的算法来学习,例如蒙特卡罗树搜索、 Q 学习、 SARSA 等。这些算法都有各自的优缺点,在不同的应用场景下选择合适的算法是非常重要的。

总之,强化学习是一种让计算机通过不断尝试来找到最优行动策略的方法,它在很多领域都有广泛的应用,例如机器人控制、博弈论等。

比较流行的强化学习模型

DQN(深度强化学习)
DDQN(双倍 DQN)
DDPG(深度强化学习确定策略梯度)
A2C(同步强化学习的连续动作值)
PPO(有效的策略梯度)
TRPO(无模型正则化策略梯度)
SAC(确定性策略梯度)
D4PG(分布式 DDPG)
D3PG(分布式 DDPG with Delay)
TD3(模仿估算器梯度计算)
MADDPG(多智能体分布式 DDPG)
HER(层次化模拟)
CER(优化层次化模拟)
QMIX(混合多智能体深度强化学习)
COMA(协作多智能体)
ICM(内在奖励机制)
UNREAL(模仿器深度强化学习)
A3C(异步动作值计算)
DQN+(深度强化学习的加强版)
GAE(基于平滑估计的反馈权衡)
ACER(优化异步策略梯度)
Ape-X DQN(分布式 DQN)
ACKTR(注意力的同步强化学习的连续动作值)
DDPG+(加强版 DDPG)
DQfD(基于目标的 DQN)

为了方便学习,我这里补充了结构模型的实现案例:
Q-learning: https://github.com/dennybritz/reinforcement-learning/tree/master/lib/envs/cliff_walking
SARSA: https://github.com/dennybritz/reinforcement-learning/tree/master/lib/envs/windy_gridworld
Deep Q-Network (DQN): https://github.com/openai/gym/tree/master/gym/envs/atari/atari_env.py
Asynchronous Advantage Actor-critic (A3C): https://github.com/openai/universe-starter-agent
Trust Region Policy Optimization (TRPO): https://github.com/openai/baselines/tree/master/baselines/trpo_mpi
Proximal Policy Optimization (PPO): https://github.com/openai/baselines/tree/master/baselines/ppo1
DDPG: https://github.com/openai/baselines/tree/master/baselines/ddpg
Soft Actor-critic (SAC): https://github.com/pranz24/pytorch-soft-actor-critic
TD3: https://github.com/sfujim/TD3
QT-Opt: https://github.com/google-research/qt-opt
Hill-Climbing: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/hill_climbing
Evolution Strategies: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/evolution_strategy
Genetic Algorithm: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/genetic_algorithm
Particle Swarm Optimization: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/particle_swarm_optimization
Cross-Entropy Method: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/cross_entropy_method
Stochastic Gradient Descent: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/stochastic_gradient_descent
Gradient Descent: https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/gradient_descent
Natural Evolution Strategies (NES): https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/natural_evolution_strategies
Covariance Matrix Adaptation Evolution Strategy (CMA-ES): https://github.com/ShangtongZhang/reinforcement-learning-an-introduction/tree/master/cma_es

也还有其他地址,重复的请仔细甄别:
Q-Learning: https://github.com/openai/gym/blob/master/examples/qlearn.py
Deep Q-Learning (DQN): https://github.com/openai/gym/blob/master/examples/dqn.py
SARSA: https://github.com/openai/gym/blob/master/examples/sarsa.py
Deep SARSA (DSARSA): https://github.com/openai/gym/blob/master/examples/dsarsa.py
A3C: https://github.com/openai/gym/blob/master/examples/a3c.py
DDPG: https://github.com/openai/gym/blob/master/examples/ddpg.py
TRPO: https://github.com/openai/gym/blob/master/examples/trpo.py
PPO: https://github.com/openai/gym/blob/master/examples/ppo.py
DQN + prioritized experience replay: https://github.com/openai/gym/blob/master/examples/dqn_prioritized_replay.py
DQN + Dueling architecture: https://github.com/openai/gym/blob/master/examples/dqn_dueling.py
DQN + Distributional RL: https://github.com/openai/gym/blob/master/examples/dqn_distributional.py
A2C + GAE: https://github.com/openai/gym/blob/master/examples/a2c_gae.py
A2C + continuous control: https://github.com/openai/gym/blob/master/examples/a2c_continuous.py
TD3: https://github.com/openai/gym/blob/master/examples/td3.py
SAC: https://github.com/openai/gym/blob/master/examples/sac.py
DQN + PER + Noisy Nets: https://github.com/openai/gym/blob/master/examples/dqn_noisy_net.py
DDPG + HER: https://github.com/openai/gym/blob/master/examples/ddpg_her.py
DQN + fixed Q-targets: https://github.com/openai/gym/blob/master/examples/dqn_fixed_q_targets.py
SAC + continuous control: https://github.com/openai/gym/blob/master/examples/sac_continuous.py
DDPG + PER: https://github.com/openai/gym/blob/master/examples/ddpg_per.py
DQN (Deep Q Network) 和 C51 (Categorical DQN) 是两种强化学习算法,你可以在以下地址找到它们的代码:
DQN:https://github.com/openai/baselines/tree/master/baselines/deepq
C51:https://github.com/openai/baselines/tree/master/baselines/deepq/categorical
你还可以在 OpenAI 官网的论文页面中找到 DQN 和 C51 的论文:

DQN:https://openai.com/blog/human-level-control-through-deep-reinforcement-learning/
C51:https://arxiv.org/abs/1707.06887

猜你喜欢

转载自blog.csdn.net/weixin_41194129/article/details/128572111