Aimbot-PPO/Aimbot-PPO-Python/Pytorch/GAIL.py

import os
import argparse
import wandb
import time
import numpy as np
import random
import uuid
import torch
import torch.nn as nn
import torch.optim as optim

from AimbotEnv import Aimbot
from tqdm import tqdm
from enum import Enum
from torch.distributions.normal import Normal
from torch.distributions.categorical import Categorical
from distutils.util import strtobool
from torch.utils.tensorboard import SummaryWriter
from mlagents_envs.environment import UnityEnvironment
from mlagents_envs.side_channel.side_channel import (
    SideChannel,
    IncomingMessage,
    OutgoingMessage,
)
from typing import List
from GAILMem import GAILMem

bestReward = 0

useCUDA = True
DEFAULT_SEED = 9331
ENV_PATH = "../Build/Build-ParallelEnv-Target-OffPolicy-SingleStack-SideChannel-EndReward-Easy/Aimbot-ParallelEnv"
EXPERT_PATH = "NAN"
SIDE_CHANNEL_UUID = uuid.UUID("8bbfb62a-99b4-457c-879d-b78b69066b5e")
WAND_ENTITY = "koha9"
WORKER_ID = 2
BASE_PORT = 1001

# max round steps per agent is 2500/Decision_period, 25 seconds
# !!!check every parameters before run!!!

TOTAL_STEPS = 6750000
BATCH_SIZE = 512
MAX_TRAINNING_DATASETS = 3000
DECISION_PERIOD = 1
LEARNING_RATE = 1e-3
GAMMA = 0.99
GAE_LAMBDA = 0.95
EPOCHS = 4
CLIP_COEF = 0.1
POLICY_COEF = 1.0
ENTROPY_COEF = 0.01
CRITIC_COEF = 0.5
TARGET_LEARNING_RATE = 1e-5
DSCM_ENHANCED = 1

ANNEAL_LEARNING_RATE = True
CLIP_VLOSS = True
NORM_ADV = True
TRAIN = True

WANDB_TACK = True
LOAD_DIR = None
# LOAD_DIR = "../PPO-Model/Aimbot-target-last.pt"

# public data
class Targets(Enum):
    Free = 0
    Go = 1
    Attack = 2
    Defence = 3
    Num = 4


BASE_WINREWARD = 999
BASE_LOSEREWARD = -999
TARGETNUM = 4
ENV_TIMELIMIT = 30
RESULT_BROADCAST_RATIO = 2 / ENV_TIMELIMIT
TotalRounds = {"Free": 0, "Go": 0, "Attack": 0}
WinRounds = {"Free": 0, "Go": 0, "Attack": 0}
EPS = 1e-8

# !!!SPECIAL PARAMETERS!!!
# change it while program is finished
using_targets_num = 3


def parse_args():
    # fmt: off
    # pytorch and environment parameters
    parser = argparse.ArgumentParser()
    parser.add_argument("--seed", type=int, default=DEFAULT_SEED,
                        help="seed of the experiment")
    parser.add_argument("--path", type=str, default=ENV_PATH,
                        help="enviroment path")
    parser.add_argument("--workerID", type=int, default=WORKER_ID,
                        help="unity worker ID")
    parser.add_argument("--baseport", type=int, default=BASE_PORT,
                        help="port to connect to Unity environment")
    parser.add_argument("--lr", type=float, default=LEARNING_RATE,
                        help="the learning rate of optimizer")
    parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
                        help="if toggled, cuda will be enabled by default")
    parser.add_argument("--total-timesteps", type=int, default=TOTAL_STEPS,
                        help="total timesteps of the experiments")

    # model parameters
    parser.add_argument("--train",type=lambda x: bool(strtobool(x)), default=TRAIN, nargs="?", const=True,
                        help="Train Model or not")
    parser.add_argument("--datasetSize", type=int, default=MAX_TRAINNING_DATASETS,
                        help="training dataset size,start training while dataset collect enough data")
    parser.add_argument("--minibatchSize", type=int, default=BATCH_SIZE,
                        help="nimi batch size")
    parser.add_argument("--epochs", type=int, default=EPOCHS,
                        help="the K epochs to update the policy")
    parser.add_argument("--annealLR", type=lambda x: bool(strtobool(x)), default=ANNEAL_LEARNING_RATE, nargs="?", const=True,
                        help="Toggle learning rate annealing for policy and value networks")
    parser.add_argument("--wandb-track", type=lambda x: bool(strtobool(x)), default=WANDB_TACK, nargs="?", const=True,
                        help="track on the wandb")
    parser.add_argument("--wandb-entity", type=str, default=WAND_ENTITY,
                        help="the entity (team) of wandb's project")
    parser.add_argument("--load-dir", type=str, default=LOAD_DIR,
                        help="load model directory")
    parser.add_argument("--decision-period", type=int, default=DECISION_PERIOD,
                        help="the number of steps to run in each environment per policy rollout")

    # GAE loss
    parser.add_argument("--gae", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
                        help="Use GAE for advantage computation")
    parser.add_argument("--norm-adv", type=lambda x: bool(strtobool(x)), default=NORM_ADV, nargs="?", const=True,
                        help="Toggles advantages normalization")
    parser.add_argument("--gamma", type=float, default=GAMMA,
                        help="the discount factor gamma")
    parser.add_argument("--gaeLambda", type=float, default=GAE_LAMBDA,
                        help="the lambda for the general advantage estimation")
    parser.add_argument("--clip-coef", type=float, default=CLIP_COEF,
                        help="the surrogate clipping coefficient")
    parser.add_argument("--policy-coef", type=float, default=POLICY_COEF,
                        help="coefficient of the policy")
    parser.add_argument("--ent-coef", type=float, default=ENTROPY_COEF,
                        help="coefficient of the entropy")
    parser.add_argument("--critic-coef", type=float, default=CRITIC_COEF,
                        help="coefficient of the value function")
    parser.add_argument("--clip-vloss", type=lambda x: bool(strtobool(x)), default=CLIP_VLOSS, nargs="?", const=True,
                        help="Toggles whether or not to use a clipped loss for the value function, as per the paper.")
    parser.add_argument("--max-grad-norm", type=float, default=0.5,
                        help="the maximum norm for the gradient clipping")
    parser.add_argument("--target-kl", type=float, default=None,
                        help="the target KL divergence threshold")
    # fmt: on
    args = parser.parse_args()
    return args


def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
    torch.nn.init.orthogonal_(layer.weight, std)
    torch.nn.init.constant_(layer.bias, bias_const)
    return layer


class PPOAgent(nn.Module):
    def __init__(self, env: Aimbot, targetNum: int):
        super(PPOAgent, self).__init__()
        self.targetNum = targetNum
        self.discrete_size = env.unity_discrete_size
        self.discrete_shape = list(env.unity_discrete_branches)
        self.continuous_size = env.unity_continuous_size

        self.network = nn.ModuleList(
            [
                nn.Sequential(
                    layer_init(nn.Linear(np.array(env.unity_observation_shape).prod(), 300)),
                    nn.ReLU(),
                    layer_init(nn.Linear(300, 200)),
                    nn.ReLU(),
                )
                for i in range(targetNum)
            ]
        )
        self.actor_dis = nn.ModuleList(
            [layer_init(nn.Linear(200, self.discrete_size), std=0.01) for i in range(targetNum)]
        )
        self.actor_mean = nn.ModuleList(
            [layer_init(nn.Linear(200, self.continuous_size), std=0.01) for i in range(targetNum)]
        )
        self.actor_logstd = nn.ParameterList(
            [nn.Parameter(torch.zeros(1, self.continuous_size)) for i in range(targetNum)]
        )
        self.critic = nn.ModuleList(
            [layer_init(nn.Linear(200, 1), std=1) for i in range(targetNum)]
        )

    def get_value(self, state: torch.Tensor):
        targets = state[:, 0].to(torch.int32)
        hidden = torch.stack([self.network[targets[i]](state[i]) for i in range(targets.size()[0])])
        return torch.stack([self.critic[targets[i]](hidden[i]) for i in range(targets.size()[0])])

    def get_actions_value(self, state: torch.Tensor, actions=None):
        targets = state[:, 0].to(torch.int32)
        hidden = torch.stack([self.network[targets[i]](state[i]) for i in range(targets.size()[0])])

        # discrete
        # 递归targets的数量,既agent数来实现根据target不同来选用对应的输出网络计算输出
        dis_logits = torch.stack(
            [self.actor_dis[targets[i]](hidden[i]) for i in range(targets.size()[0])]
        )
        split_logits = torch.split(dis_logits, self.discrete_shape, dim=1)
        multi_categoricals = [Categorical(logits=thisLogits) for thisLogits in split_logits]
        # continuous
        actions_mean = torch.stack(
            [self.actor_mean[targets[i]](hidden[i]) for i in range(targets.size()[0])]
        )  # self.actor_mean(hidden)
        # action_logstd = torch.stack([self.actor_logstd[targets[i]].expand_as(actions_mean) for i in range(targets.size()[0])]) # self.actor_logstd.expand_as(actions_mean)
        # print(action_logstd)
        action_std = torch.squeeze(
            torch.stack(
                [torch.exp(self.actor_logstd[targets[i]]) for i in range(targets.size()[0])]
            ),
            dim=-1,
        )  # torch.exp(action_logstd)
        con_probs = Normal(actions_mean, action_std)
        # critic
        criticV = torch.stack(
            [self.critic[targets[i]](hidden[i]) for i in range(targets.size()[0])]
        )

        if actions is None:
            if TRAIN:
                # select actions base on probability distribution model
                disAct = torch.stack([ctgr.sample() for ctgr in multi_categoricals])
                conAct = con_probs.sample()
                actions = torch.cat([disAct.T, conAct], dim=1)
            else:
                # select actions base on best probability distribution
                disAct = torch.stack([torch.argmax(logit, dim=1) for logit in split_logits])
                conAct = actions_mean
                actions = torch.cat([disAct.T, conAct], dim=1)
        else:
            disAct = actions[:, 0 : env.unity_discrete_type].T
            conAct = actions[:, env.unity_discrete_type :]
        dis_log_prob = torch.stack(
            [ctgr.log_prob(act) for act, ctgr in zip(disAct, multi_categoricals)]
        )
        dis_entropy = torch.stack([ctgr.entropy() for ctgr in multi_categoricals])

        return (
            actions,
            dis_log_prob.sum(0),
            dis_entropy.sum(0),
            con_probs.log_prob(conAct).sum(1),
            con_probs.entropy().sum(1),
            criticV,
        )


def GAE(agent, args, rewards, dones, values, next_obs, next_done):
    # GAE
    with torch.no_grad():
        next_value = agent.get_value(next_obs).reshape(1, -1)
        data_size = rewards.size()[0]
        if args.gae:
            advantages = torch.zeros_like(rewards).to(device)
            lastgaelam = 0
            for t in reversed(range(data_size)):
                if t == data_size - 1:
                    nextnonterminal = 1.0 - next_done
                    nextvalues = next_value
                else:
                    nextnonterminal = 1.0 - dones[t + 1]
                    nextvalues = values[t + 1]
                delta = rewards[t] + args.gamma * nextvalues * nextnonterminal - values[t]
                advantages[t] = lastgaelam = (
                    delta + args.gamma * args.gaeLambda * nextnonterminal * lastgaelam
                )
            returns = advantages + values
        else:
            returns = torch.zeros_like(rewards).to(device)
            for t in reversed(range(data_size)):
                if t == data_size - 1:
                    nextnonterminal = 1.0 - next_done
                    next_return = next_value
                else:
                    nextnonterminal = 1.0 - dones[t + 1]
                    next_return = returns[t + 1]
                returns[t] = rewards[t] + args.gamma * nextnonterminal * next_return
            advantages = returns - values
    return advantages, returns


class AimbotSideChannel(SideChannel):
    def __init__(self, channel_id: uuid.UUID) -> None:
        super().__init__(channel_id)

    def on_message_received(self, msg: IncomingMessage) -> None:
        """
        Note: We must implement this method of the SideChannel interface to
        receive messages from Unity
        """
        thisMessage = msg.read_string()
        # print(thisMessage)
        thisResult = thisMessage.split("|")
        if thisResult[0] == "result":
            TotalRounds[thisResult[1]] += 1
            if thisResult[2] == "Win":
                WinRounds[thisResult[1]] += 1
            # print(TotalRounds)
            # print(WinRounds)
        elif thisResult[0] == "Error":
            print(thisMessage)

    # 发送函数
    def send_string(self, data: str) -> None:
        # send a string toC#
        msg = OutgoingMessage()
        msg.write_string(data)
        super().queue_message_to_send(msg)

    def send_bool(self, data: bool) -> None:
        msg = OutgoingMessage()
        msg.write_bool(data)
        super().queue_message_to_send(msg)

    def send_int(self, data: int) -> None:
        msg = OutgoingMessage()
        msg.write_int32(data)
        super().queue_message_to_send(msg)

    def send_float(self, data: float) -> None:
        msg = OutgoingMessage()
        msg.write_float32(data)
        super().queue_message_to_send(msg)

    def send_float_list(self, data: List[float]) -> None:
        msg = OutgoingMessage()
        msg.write_float32_list(data)
        super().queue_message_to_send(msg)


def broadCastEndReward(rewardBF: list, remainTime: float):
    thisRewardBF = rewardBF
    if rewardBF[-1] <= -500:
        # print("Lose DO NOT BROAD CAST",rewardBF[-1])
        thisRewardBF[-1] = rewardBF[-1] - BASE_LOSEREWARD
        thisRewardBF = (np.asarray(thisRewardBF)).tolist()
    elif rewardBF[-1] >= 500:
        # print("Win! Broadcast reward!",rewardBF[-1])
        thisRewardBF[-1] = rewardBF[-1] - BASE_WINREWARD
        thisRewardBF = (np.asarray(thisRewardBF) + (remainTime * RESULT_BROADCAST_RATIO)).tolist()
    else:
        print("!!!!!DIDNT GET RESULT REWARD!!!!!!", rewardBF[-1])
    return torch.Tensor(thisRewardBF).to(device)


class Discriminator(nn.Module):
    def __init__(self, env: Aimbot, targetNum: int):
        super(Discriminator, self).__init__()
        self.targetNum = targetNum
        self.trajectory_size = env.unity_observation_shape + env.unity_action_size

        self.network = nn.ModuleList(
            [
                nn.Sequential(
                    layer_init(nn.Linear(np.array(self.trajectory_size).prod(), 300)),
                    nn.ReLU(),
                    layer_init(nn.Linear(300, 200)),
                    nn.ReLU(),
                )
                for i in range(targetNum)
            ]
        )
        self.output = nn.ModuleList(
            [nn.Relu(layer_init(nn.Linear(200, 1), std=0.01)) for i in range(targetNum)]
        )

    def get_value(self, state: torch.Tensor):
        targets = state[:, 0].to(torch.int32)
        hidden = torch.stack([self.network[targets[i]](state[i]) for i in range(targets.size()[0])])
        return torch.stack([self.output[targets[i]](hidden[i]) for i in range(targets.size()[0])])


if __name__ == "__main__":
    args = parse_args()
    random.seed(DEFAULT_SEED)
    np.random.seed(DEFAULT_SEED)
    torch.manual_seed(DEFAULT_SEED)

    device = torch.device("cuda" if torch.cuda.is_available() and useCUDA else "cpu")

    aimBotsideChannel = AimbotSideChannel(SIDE_CHANNEL_UUID)
    env = Aimbot(
        envPath=ENV_PATH, workerID=WORKER_ID, basePort=BASE_PORT, side_channels=[aimBotsideChannel]
    )
    if LOAD_DIR is None:
        agent = PPOAgent(env, TARGETNUM).to(device)
        discriminator = Discriminator(env, TARGETNUM).to(device)
    else:
        print("NOT AVALABLE")
        agent = torch.load(LOAD_DIR)
        print("Load Agent", LOAD_DIR)
        print(agent.eval())

    # Optimizers
    PPOoptimizer = optim.Adam(agent.parameters(), lr=LEARNING_RATE, eps=1e-5)
    DSCMoptimizer = optim.Adam(discriminator.parameters(), lr=LEARNING_RATE, eps=1e-5)

    # Tensorboard and WandB Recorder
    game_name = "Aimbot_GAIL"
    game_type = "ORG"
    run_name = f"{game_name}_{game_type}_{DEFAULT_SEED}_{int(time.time())}"
    if WANDB_TACK:
        wandb.init(
            project=game_name,
            entity=WAND_ENTITY,
            sync_tensorboard=True,
            config=vars(args),
            name=run_name,
            monitor_gym=True,
            save_code=True,
        )
    writer = SummaryWriter(f"runs/{run_name}")
    writer.add_text(
        "hyperparameters",
        "|param|value|\n|-|-|\n%s"
        % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
    )

    agentMem = GAILMem()
    expertMem = GAILMem()
    expertMem.loadMemFile(EXPERT_PATH)

    # start the game
    total_update_step = using_targets_num * args.total_timesteps // args.datasetSize
    target_steps = [0 for i in range(TARGETNUM)]
    start_time = time.time()
    state, _, done = env.reset()
    # Trajectory Buffer
    ob_bf = [[] for i in range(env.unity_agent_num)]
    act_bf = [[] for i in range(env.unity_agent_num)]
    dis_logprobs_bf = [[] for i in range(env.unity_agent_num)]
    con_logprobs_bf = [[] for i in range(env.unity_agent_num)]
    rewards_bf = [[] for i in range(env.unity_agent_num)]
    dones_bf = [[] for i in range(env.unity_agent_num)]
    values_bf = [[] for i in range(env.unity_agent_num)]

    # initialize empty training datasets
    obs = [
        torch.tensor([]).to(device) for i in range(TARGETNUM)
    ]  # (TARGETNUM,n,env.unity_observation_size)
    actions = [
        torch.tensor([]).to(device) for i in range(TARGETNUM)
    ]  # (TARGETNUM,n,env.unity_action_size)
    dis_logprobs = [torch.tensor([]).to(device) for i in range(TARGETNUM)]  # (TARGETNUM,n,1)
    con_logprobs = [torch.tensor([]).to(device) for i in range(TARGETNUM)]  # (TARGETNUM,n,1)
    rewards = [torch.tensor([]).to(device) for i in range(TARGETNUM)]  # (TARGETNUM,n,1)
    values = [torch.tensor([]).to(device) for i in range(TARGETNUM)]  # (TARGETNUM,n,1)
    dones = [torch.tensor([]).to(device) for i in range(TARGETNUM)]

    next_state_buffer = [[] for i in range(TARGETNUM)]
    next_done_buffer = [[] for i in range(TARGETNUM)]
    for total_steps in range(total_update_step):
        print("new episode")
        trainQueue = []
        while True:
            with torch.no_grad():
                # predict actions
                action, dis_logprob, _, con_logprob, _, value = agent.get_actions_value(
                    torch.Tensor(state).to(device)
                )
                value = value.flatten()

            # variable from GPU to CPU
            action_cpu = action.cpu().numpy()
            dis_logprob_cpu = dis_logprob.cpu().numpy()
            con_logprob_cpu = con_logprob.cpu().numpy()
            value_cpu = value.cpu().numpy()

            # Environment step
            next_state, reward, next_done = env.step(action_cpu)

            # save mem
            for i in range(env.unity_agent_num):
                # save memories to buffers
                ob_bf[i].append(state[i])
                act_bf[i].append(action_cpu[i])
                dis_logprobs_bf[i].append(dis_logprob_cpu[i])
                con_logprobs_bf[i].append(con_logprob_cpu[i])
                rewards_bf[i].append(reward[i])
                dones_bf[i].append(done[i])
                values_bf[i].append(value_cpu[i])
                if next_done[i] == True:
                    # finished a round, send finished memories to training datasets
                    thisTarget = int(state[i, 0])
                    next_state_buffer[thisTarget] = next_state
                    next_done_buffer[thisTarget] = next_done
                    obs[thisTarget] = torch.cat(
                        (obs[thisTarget], torch.tensor(ob_bf[i]).to(device)), 0
                    )
                    actions[thisTarget] = torch.cat(
                        (actions[thisTarget], torch.tensor(act_bf[i]).to(device)), 0
                    )
                    dis_logprobs[thisTarget] = torch.cat(
                        (dis_logprobs[thisTarget], torch.tensor(dis_logprobs_bf[i]).to(device)), 0
                    )
                    con_logprobs[thisTarget] = torch.cat(
                        (con_logprobs[thisTarget], torch.tensor(con_logprobs_bf[i]).to(device)), 0
                    )
                    values[thisTarget] = torch.cat(
                        (values[thisTarget], torch.tensor(values_bf[i]).to(device)), 0
                    )
                    dones[thisTarget] = torch.cat(
                        (dones[thisTarget], torch.tensor(dones_bf[i]).to(device)), 0
                    )
            for i in range(TARGETNUM):
                if obs[i].size()[0] >= args.datasetSize:
                    # start train NN
                    trainQueue.append(i)
            if len(trainQueue) > 0:
                break
            state, done = next_state, next_done

        if args.train:
            meanRewardList = []
            for thisT in trainQueue:
                target_steps[thisT] += 1
                # get agent training datasets
                b_obs = obs[thisT].reshape((-1,) + env.unity_observation_shape)
                b_actions = actions[thisT].reshape((-1,) + (env.unity_action_size,))
                b_dis_logprobs = dis_logprobs[thisT].reshape(-1)
                b_con_logprobs = con_logprobs[thisT].reshape(-1)
                b_values = values[thisT].reshape(-1)
                b_dones = dones[thisT].reshape(-1)
                dataNum = b_obs[thisT].size()[0]
                # get expert training datasets from GAILMem
                exp_obs, _, exp_actions, _, _ = expertMem.getRandomSample(dataNum, thisT)

                # trajectory
                agent_trajectory = torch.cat((b_obs, b_actions), dim=1)
                exp_trajectory = torch.cat((exp_obs, exp_actions), dim=1)
                
                # discriminator ACC
                with torch.no_grad():
                    thisDSCM_agent_acc = torch.mean(discriminator.get_value(agent_trajectory))
                    thisDSCM_expt_acc = torch.mean(discriminator.get_value(exp_trajectory))
                # train discriminator
                b_inds = np.arange(dataNum)
                for epoch in range(args.epoch):
                    np.random.shuffle(b_inds)
                    # train discriminator
                    for start in range(0, dataNum, args.minibatchSize):
                        end = start + args.minibatchSize
                        mb_inds = b_inds[start:end]
                        exp_value = discriminator.get_value(exp_trajectory[mb_inds])
                        agent_value = discriminator.get_value(agent_trajectory[mb_inds])
                        # DSCM loss function
                        exp_loss = torch.log(1.0 - exp_value + EPS)
                        agent_loss = torch.log(agent_value + EPS)
                        loss = exp_loss + agent_loss
                        # DSCM backward
                        DSCMoptimizer.zero_grad()
                        loss.backward()
                        nn.utils.clip_grad_norm_(discriminator.parameters(), args.max_grad_norm)
                        DSCMoptimizer.step()
                # get discriminator reward
                with torch.no_grad():
                    DSCMReward = discriminator.get_value(agent_trajectory) * DSCM_ENHANCED
                advantages, rts = GAE(
                    agent,
                    args,
                    DSCMReward,
                    b_dones,
                    b_values,
                    next_state_buffer[thisTarget],
                    next_done_buffer[thisTarget],
                )
                b_advantages = advantages.reshape(-1)
                b_returns = rts.reshape(-1)

                # train PPO agent
                for epoch in range(args.epoch):
                    np.random.shuffle(b_inds)
                    for start in range(0, dataNum, args.minibatchSize):
                        end = start + args.minibatchSize
                        mb_inds = b_inds[start:end]
                        mb_advantages = b_advantages[mb_inds]
                        # normalize advantages
                        if args.norm_adv:
                            mb_advantages = (mb_advantages - mb_advantages.mean()) / (
                                mb_advantages.std() + 1e-8
                            )
                        (
                            _,
                            new_dis_logprob,
                            dis_entropy,
                            new_con_logprob,
                            con_entropy,
                            newvalue,
                        ) = agent.get_actions_value(b_obs[mb_inds], b_actions[mb_inds])
                        # discrete ratio
                        dis_logratio = new_dis_logprob - b_dis_logprobs[mb_inds]
                        dis_ratio = dis_logratio.exp()
                        # continuous ratio
                        con_logratio = new_con_logprob - b_con_logprobs[mb_inds]
                        con_ratio = con_logratio.exp()
                        # discrete Policy loss
                        dis_pg_loss_orig = -mb_advantages * dis_ratio
                        dis_pg_loss_clip = -mb_advantages * torch.clamp(
                            dis_ratio, 1 - args.clip_coef, 1 + args.clip_coef
                        )
                        dis_pg_loss = torch.max(dis_pg_loss_orig, dis_pg_loss_clip).mean()
                        # continuous Policy loss
                        con_pg_loss_orig = -mb_advantages * con_ratio
                        con_pg_loss_clip = -mb_advantages * torch.clamp(
                            con_ratio, 1 - args.clip_coef, 1 + args.clip_coef
                        )
                        con_pg_loss = torch.max(con_pg_loss_orig, con_pg_loss_clip).mean()
                        # Value loss
                        newvalue = newvalue.view(-1)
                        if args.clip_vloss:
                            v_loss_unclipped = (newvalue - b_returns[mb_inds]) ** 2
                            v_clipped = b_values[mb_inds] + torch.clamp(
                                newvalue - b_values[mb_inds],
                                -args.clip_coef,
                                args.clip_coef,
                            )
                            v_loss_clipped = (v_clipped - b_returns[mb_inds]) ** 2
                            v_loss_max = torch.max(v_loss_unclipped, v_loss_clipped)
                            v_loss = 0.5 * v_loss_max.mean()
                        else:
                            v_loss = 0.5 * ((newvalue - b_returns[mb_inds]) ** 2).mean()
                        # total loss
                        entropy_loss = dis_entropy.mean() + con_entropy.mean()
                        loss = (
                            dis_pg_loss * args.policy_coef
                            + con_pg_loss * args.policy_coef
                            - entropy_loss * args.ent_coef
                            + v_loss * args.critic_coef
                        )
                        PPOoptimizer.zero_grad()
                        loss.backward()
                        # Clips gradient norm of an iterable of parameters.
                        nn.utils.clip_grad_norm_(agent.parameters(), args.max_grad_norm)
                        PPOoptimizer.step()
                # record mean reward before clear history
                targetRewardMean = np.mean(rewards[thisT].to("cpu").detach().numpy().copy())
                meanRewardList.append(targetRewardMean)
                targetName = Targets(thisT).name

                # clear this target trainning set buffer
                obs[thisT] = torch.tensor([]).to(device)
                actions[thisT] = torch.tensor([]).to(device)
                dis_logprobs[thisT] = torch.tensor([]).to(device)
                con_logprobs[thisT] = torch.tensor([]).to(device)
                rewards[thisT] = torch.tensor([]).to(device)
                values[thisT] = torch.tensor([]).to(device)
                dones[thisT] = torch.tensor([]).to(device)

                # record rewards for plotting purposes
                writer.add_scalar(f"Target{targetName}/value_loss", v_loss.item(), target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/dis_policy_loss", dis_pg_loss.item(), target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/con_policy_loss", con_pg_loss.item(), target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/total_loss", loss.item(), target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/entropy_loss", entropy_loss.item(), target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/Reward", targetRewardMean, target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/WinRatio", WinRounds[targetName]/TotalRounds[targetName], target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/Discriminator_EXP_ACC", thisDSCM_expt_acc,target_steps[thisT])
                writer.add_scalar(f"Target{targetName}/Discriminator_Agent_ACC", thisDSCM_agent_acc,target_steps[thisT])
                print(f"episode over Target{targetName} mean reward:", targetRewardMean)
            TotalRewardMean = np.mean(meanRewardList)
            writer.add_scalar("GlobalCharts/TotalRewardMean", TotalRewardMean, total_steps)
            writer.add_scalar("GlobalCharts/learning_rate", PPOoptimizer.param_groups[0]["lr"], total_steps)
            # New Record!
            if TotalRewardMean > bestReward:
                bestReward = targetRewardMean
                saveDir = "../PPO-Model/Hybrid-MNN-500-300" + str(TotalRewardMean) + ".pt"
                torch.save(agent, saveDir)

        # train PPO
        # record