实时精准控制机械臂AGIagent

CreateAMind

发布于 2023-11-30 13:23:08

1410

发布于 2023-11-30 13:23:08

文章被收录于专栏：CreateAMind

AGI agent 通过贝叶斯推理自由能目标函数，进行实时推理并完美解决pendulum任务环境；追踪目标能力极强

先简单看一下强化学习的效果

强化学习的Pendulum 效果：

这里使用juliareinforcementlearning库；epoch步数高达5000步

functionRL.Experiment(    ::Val{:JuliaRL},    ::Val{:BasicDQN},    ::Val{:PendulumDiscrete},    ::Nothing;    seed =123, )    rng = StableRNG(seed)    env = PendulumEnv(continuous =false, max_steps =5000, rng = rng)    ns, na = length(state(env)), length(action_space(env))    agent = Agent(        policy = QBasedPolicy(            learner = BasicDQNLearner(                approximator = NeuralNetworkApproximator(                    model = Chain(                        Dense(ns,64, relu; init = glorot_uniform(rng)),                        Dense(64,64, relu; init = glorot_uniform(rng)),                        Dense(64, na; init = glorot_uniform(rng)),                    ) |> gpu,                    optimizer = ADAM(),                ),                batch_size =32,                min_replay_history =100,                loss_func = huber_loss,                rng = rng,            ),            explorer = EpsilonGreedyExplorer(                kind = :exp,                ϵ_stable =0.01,                decay_steps =500,                rng = rng,            ),        ),        trajectory = CircularArraySARTTrajectory(            capacity =5_000,            state =Vector{Float32} => (ns,),        ),    )    stop_condition = StopAfterStep(50_000, is_show_progress=!haskey(ENV,"CI"))    hook = TotalRewardPerEpisode()    Experiment(agent, env, stop_condition, hook,"")end