src/takeaway.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ORLA: learning to play Takeaway\n",
    "[_Keepaway_](https://www.cs.utexas.edu/~AustinVilla/sim/keepaway/) is a soccer-like game in which a team of _keepers_ and a team of _takers_ compete for the possession of the ball. The game always starts with one of the keepers having possession of the ball. [_Takeaway_](https://arxiv.org/abs/2004.07928) is the name that Keepaway receives when the takers are the only learning agents and the keepers follow a fixed policy. Our Takeaway matches are played with 4 keepers and 3 Takers on a 40x40 field (unless stated otherwise). Here is a screenshot showing what a newly generated Takeaway match looks like:\n",
    "<center><img src=\"../imgs/keepaway_screenshot.png\" width=\"300\"></center>\n",
    "\n",
    "To solve the takeaway task, an expert proposes the following arguments (in brackets, the corresponding naming in our paper): TakleBallX ($TB_i$), MinAngleX,Y ($A_ij$), MinDistX,Y ($C_i,j$), OpenKeeperX,Y ($O_ij$), and FarKeeperX,Y ($F_ij$). See paper for further details.\n",
    "\n",
    "This notebook explains how ORLA can be used to decode both a local ($\\texttt{M}_\\text{loc}$) and global ($\\texttt{M}_\\text{glob}$) ranking."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Setting things up\n",
    "### Imports"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Using device: cpu\n"
     ]
    }
   ],
   "source": [
    "from argumentation import utils as argm\n",
    "from environments.takeaway import utils as tkwy\n",
    "from environments.takeaway.takeaway import Takeaway\n",
    "from agents.orla import ORLABaseline\n",
    "import utils\n",
    "\n",
    "import numpy as np\n",
    "import torch\n",
    "from tqdm.notebook import tqdm\n",
    "from utils import plot_returns\n",
    "\n",
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "print('Using device:', device)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Connecting with RoboCup\n",
    "\n",
    "Takeaway is implemented on [rcssserver](https://github.com/omcandido/rcssserver_orla) and uses the [keepaway library](https://github.com/omcandido/keepaway_orla) as an interface with the learning agents. Here, ORLA assumes that takewaway runs on a WSL2 instance and connects to it via sockets:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "send_host = tkwy.get_host('send')\n",
    "send_port = 7000\n",
    "recv_host = tkwy.get_host('recv')\n",
    "recv_port = 7001"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Loading the expert arguments"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "There are 39 expert arguments: ['TackleBall1', 'OpenKeeper1,2', 'FarKeeper1,2', 'MinAngle1,2', 'MinDist1,2', 'OpenKeeper1,3', 'FarKeeper1,3', 'MinAngle1,3', 'MinDist1,3', 'OpenKeeper1,4', 'FarKeeper1,4', 'MinAngle1,4', 'MinDist1,4', 'TackleBall2', 'OpenKeeper2,2', 'FarKeeper2,2', 'MinAngle2,2', 'MinDist2,2', 'OpenKeeper2,3', 'FarKeeper2,3', 'MinAngle2,3', 'MinDist2,3', 'OpenKeeper2,4', 'FarKeeper2,4', 'MinAngle2,4', 'MinDist2,4', 'TackleBall3', 'OpenKeeper3,2', 'FarKeeper3,2', 'MinAngle3,2', 'MinDist3,2', 'OpenKeeper3,3', 'FarKeeper3,3', 'MinAngle3,3', 'MinDist3,3', 'OpenKeeper3,4', 'FarKeeper3,4', 'MinAngle3,4', 'MinDist3,4']\n"
     ]
    }
   ],
   "source": [
    "argument_actions = tkwy.arg_actions\n",
    "arguments = list(argument_actions.keys())\n",
    "\n",
    "print(\"There are {} expert arguments: {}\".format(len(arguments), arguments))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### The environment\n",
    "The environment evaluates the ranking decoded by ORLA by using it as a reasoning engine by the takers on a Takeaway match.\n",
    "The variable _ranking_path_ contains the file with the ranking decoded by ORLA, which is updated before every match and directly read by the takers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "ranking_path = r\"\\\\wsl.localhost\\Ubuntu16\\home\\candido\\keepaway_orla\\player\\orla\\ordering.txt\"\n",
    "env = Takeaway(\n",
    "    argument_actions,\n",
    "    send_host,\n",
    "    send_port,\n",
    "    recv_host,\n",
    "    recv_port,\n",
    "    ranking_path\n",
    ")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Experiment 1"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### The agent: $\\mathtt{M}_\\text{loc}$\n",
    "We call $\\mathtt{M}_\\text{loc}$ to the ORLA model resulting from learning a strict ranking over the expert arguments. Note that our learning model consists of two sets of parameters: $\\alpha^\\theta$ (the preference function, in this case, the NN) and $\\alpha^w$ (the baseline). This is the agent of the RL pipeline: "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "alpha_th = 5e-5\n",
    "alpha_w = utils.estimate_alpha_w(len(arguments), 1e5)\n",
    "\n",
    "mode = utils.Mode.STRICT\n",
    "orla_loc = ORLABaseline(arguments,alpha_th, alpha_w, device, mode)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "([['OpenKeeper2,2'],\n",
       "  ['MinDist2,2'],\n",
       "  ['FarKeeper3,3'],\n",
       "  ['OpenKeeper3,2'],\n",
       "  ['OpenKeeper2,3'],\n",
       "  ['OpenKeeper1,2'],\n",
       "  ['MinAngle2,2'],\n",
       "  ['MinAngle3,4'],\n",
       "  ['OpenKeeper2,4'],\n",
       "  ['TackleBall3'],\n",
       "  ['FarKeeper1,3'],\n",
       "  ['OpenKeeper1,3'],\n",
       "  ['TackleBall1'],\n",
       "  ['OpenKeeper3,4'],\n",
       "  ['MinAngle1,3'],\n",
       "  ['FarKeeper3,4'],\n",
       "  ['MinDist1,2'],\n",
       "  ['MinDist1,4'],\n",
       "  ['FarKeeper3,2'],\n",
       "  ['OpenKeeper1,4'],\n",
       "  ['MinDist1,3'],\n",
       "  ['MinAngle3,3'],\n",
       "  ['MinDist3,2'],\n",
       "  ['MinDist3,3'],\n",
       "  ['MinDist2,3'],\n",
       "  ['MinAngle3,2'],\n",
       "  ['FarKeeper2,4'],\n",
       "  ['MinDist3,4'],\n",
       "  ['MinDist2,4'],\n",
       "  ['TackleBall2'],\n",
       "  ['MinAngle1,2'],\n",
       "  ['OpenKeeper3,3'],\n",
       "  ['FarKeeper1,2'],\n",
       "  ['MinAngle1,4'],\n",
       "  ['MinAngle2,3'],\n",
       "  ['FarKeeper2,3'],\n",
       "  ['FarKeeper1,4'],\n",
       "  ['FarKeeper2,2'],\n",
       "  ['MinAngle2,4']],\n",
       " tensor([0.0282, 0.0368, 0.0241, 0.0325, 0.0309, 0.0269, 0.0247, 0.0341, 0.0364,\n",
       "         0.0336, 0.0306, 0.0460, 0.0484, 0.0350, 0.0428, 0.0454, 0.0320, 0.0500,\n",
       "         0.0314, 0.0484, 0.0513, 0.0680, 0.0609, 0.0622, 0.0658, 0.0635, 0.0960,\n",
       "         0.0968, 0.0788, 0.1061, 0.1153, 0.1589, 0.1665, 0.1516, 0.1740, 0.2402,\n",
       "         0.3712, 0.4871, 1.0000], grad_fn=<StackBackward0>))"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "orla_loc.decode_ranking()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Load the model\n",
    "Optionally, we can load a (pre)trained model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "# orla_loc.w = np.load(\"saved_models/orla_local/w.npy\")\n",
    "# orla_loc.net.load_state_dict(torch.load(\"saved_models/orla_local/model.pt\"))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Training\n",
    "We train in mini-batches.\n",
    "\n",
    "Note: it is a good idea to visualize in the beginning how the probabilities of picking first arguments 0, 13 and 26 (TackleBall arguments) evolve to check that the learning rate is not too high. I chose these arguments because they are expected to be the most important ones, according to the expert, so seeing their probabilities slowly increase is a good sign."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "ffb77cc181ec4016b370679523dfe4f6",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/4000 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "batch_size = 8\n",
    "\n",
    "returns = []\n",
    "rankings_batch = []\n",
    "probs_batch = []\n",
    "returns_batch = []\n",
    "\n",
    "t_episodes = tqdm(range(1, int(4e3)+1))\n",
    "for episode in t_episodes:\n",
    "    orla_loc.net.train()\n",
    "    ranking, prob = orla_loc.decode_ranking()\n",
    "    score = env.play(ranking)\n",
    "    returns.append(score)\n",
    "\n",
    "    rankings_batch.append(ranking)\n",
    "    probs_batch.append(prob)\n",
    "    returns_batch.append(score)\n",
    "    \n",
    "    if len(rankings_batch) == batch_size:\n",
    "        orla_loc.learn(rankings_batch, probs_batch, returns_batch)\n",
    "        rankings_batch = []\n",
    "        probs_batch = []\n",
    "        returns_batch = []\n",
    "        past = 1000\n",
    "        t_episodes.set_postfix({'avg': np.average(returns[-past:]), 'last': np.sum(returns[-batch_size:])/batch_size})\n",
    "\n",
    "    # # Print the probabilities of Tackleball_{1,2 and 3} being the first chosen argument.\n",
    "    # # It can give you an idea of how good your learning rate is:\n",
    "    # orla_loc.net.eval()\n",
    "    # ranking = []\n",
    "    # state = argm.ranking_to_matrix(ranking, args, True)\n",
    "    # state = torch.from_numpy(state).float().flatten().to(device)\n",
    "    # remaining = orla_loc.remaining_arguments(ranking)\n",
    "    # mask = orla_loc.mask_remaining(remaining)\n",
    "    # orla_loc.net.mask = torch.from_numpy(mask).bool().to(device)\n",
    "    # print(orla_loc.net(state)[0].item(), orla_loc.net(state)[13].item(), orla_loc.net(state)[26].item())\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Learning curve\n",
    "We plot the learning curve of the model by applying a rolling window (feel free to train it further to ensure convergence)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_returns(returns, window = 1000)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Save the model\n",
    "If you are done training the model, you can save it:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.save(\"saved_models/orla_local/w.npy\", orla_loc.w)\n",
    "torch.save(orla_loc.net.state_dict(), \"saved_models/orla_local/model.pt\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Analysis of the local strategy\n",
    "Let's see what strategy each of the $\\mathtt{M}_\\text{loc}$ takers has learned (note that this is model has been trained for demonstration purposes and its policy may differ from that shown in the paper):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "    T1:  [['TackleBall1'], ['MinAngle1,2'], ['MinAngle1,3'], ['OpenKeeper1,4'], ['OpenKeeper1,2'], ['OpenKeeper1,3'], ['MinDist1,3'], ['MinDist1,4'], ['FarKeeper1,4'], ['MinAngle1,4'], ['FarKeeper1,3'], ['FarKeeper1,2'], ['MinDist1,2']]\n",
      "    T2:  [['TackleBall2'], ['MinAngle2,3'], ['MinDist2,4'], ['MinAngle2,4'], ['OpenKeeper2,2'], ['OpenKeeper2,3'], ['MinAngle2,2'], ['OpenKeeper2,4'], ['FarKeeper2,3'], ['MinDist2,2'], ['FarKeeper2,4'], ['FarKeeper2,2'], ['MinDist2,3']]\n",
      "    T3:  [['TackleBall3'], ['FarKeeper3,4'], ['MinDist3,2'], ['MinAngle3,3'], ['FarKeeper3,3'], ['MinAngle3,4'], ['MinAngle3,2'], ['OpenKeeper3,3'], ['OpenKeeper3,2'], ['OpenKeeper3,4'], ['MinDist3,3'], ['MinDist3,4'], ['FarKeeper3,2']]\n"
     ]
    }
   ],
   "source": [
    "ranking, probs = orla_loc.decode_ranking(True)\n",
    "tkwy.show_ranking_by_taker(ranking)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Experiment 2\n",
    "### The agent: $\\mathtt{M}_\\text{glob}$\n",
    "We call $\\mathtt{M}_\\text{glob}$ to the ORLA model resulting from learning a non-strict ranking over the expert arguments, by grouping the average rank of the arguments of the same type."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "alpha_th = 4e-5\n",
    "alpha_w = utils.estimate_alpha_w(len(args), 1e5)\n",
    "\n",
    "MODE = utils.Mode.STRICT\n",
    "orla_glob = ORLABaseline(args,alpha_th, alpha_w, device, mode)\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Optionally, load a pre(trained) model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "# orla_glob.w = np.load(\"saved_models/orla_global/w.npy\")\n",
    "# orla_glob.net.load_state_dict(torch.load(\"saved_models/orla_global/model.pt\"))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "497552239b07456dbad9b06f6c069349",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/4000 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "batch_size = 8\n",
    "\n",
    "returns = []\n",
    "rankings_batch = []\n",
    "probs_batch = []\n",
    "returns_batch = []\n",
    "\n",
    "t_episodes = tqdm(range(1, int(4e3)+1))\n",
    "for episode in t_episodes:\n",
    "    orla_glob.net.train()\n",
    "    ranking, prob = orla_glob.decode_ranking()\n",
    "    # Now obtain the global ranking:\n",
    "    _, arg_val = tkwy.get_global_values(ranking)\n",
    "    ranking = argm.values_to_ranking(arg_val)\n",
    "    score = env.play(ranking)\n",
    "    returns.append(score)\n",
    "\n",
    "    rankings_batch.append(ranking)\n",
    "    probs_batch.append(prob)\n",
    "    returns_batch.append(score)\n",
    "    \n",
    "    if len(rankings_batch) == batch_size:\n",
    "        orla_glob.learn(rankings_batch, probs_batch, returns_batch)\n",
    "        rankings_batch = []\n",
    "        probs_batch = []\n",
    "        returns_batch = []\n",
    "        past = 1000\n",
    "        t_episodes.set_postfix({'avg': np.average(returns[-past:]), 'last': np.sum(returns[-batch_size:])/batch_size})\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Learning curve\n",
    "We plot the learning curve of the model by applying a rolling window (feel free to train it further to ensure convergence)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_returns(returns, window = 1000)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Save the model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.save(\"saved_models/orla_global/w.npy\", orla_glob.w)\n",
    "torch.save(orla_glob.net.state_dict(), \"saved_models/orla_global/model.pt\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Analysis of the global strategy\n",
    "Let's see what global strategy $\\mathtt{M}_\\text{glob}$ takers have learned (note that this is model has been trained for demonstration purposes and its policy may differ from that shown in the paper):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "    T1:  [['TackleBall1'], ['MinDist1,2', 'MinDist1,3', 'MinDist1,4'], ['MinAngle1,2', 'MinAngle1,3', 'MinAngle1,4'], ['FarKeeper1,2', 'FarKeeper1,3', 'FarKeeper1,4'], ['OpenKeeper1,2', 'OpenKeeper1,3', 'OpenKeeper1,4']]\n",
      "    T2:  [['TackleBall2'], ['MinDist2,2', 'MinDist2,3', 'MinDist2,4'], ['MinAngle2,2', 'MinAngle2,3', 'MinAngle2,4'], ['FarKeeper2,2', 'FarKeeper2,3', 'FarKeeper2,4'], ['OpenKeeper2,2', 'OpenKeeper2,3', 'OpenKeeper2,4']]\n",
      "    T3:  [['TackleBall3'], ['MinDist3,2', 'MinDist3,3', 'MinDist3,4'], ['MinAngle3,2', 'MinAngle3,3', 'MinAngle3,4'], ['FarKeeper3,2', 'FarKeeper3,3', 'FarKeeper3,4'], ['OpenKeeper3,2', 'OpenKeeper3,3', 'OpenKeeper3,4']]\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'TackleBall': 38,\n",
       " 'MinDist': 27,\n",
       " 'MinAngle': 20,\n",
       " 'FarKeeper': 14,\n",
       " 'OpenKeeper': 12}"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "ranking, probs = orla_glob.decode_ranking(True)\n",
    "glob, arg_val = tkwy.get_global_values(ranking)\n",
    "ranking = argm.values_to_ranking(arg_val)\n",
    "tkwy.show_ranking_by_taker(ranking)\n",
    "glob\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Save the decoded global ranking"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "argm.save_values(ranking_path, args, arg_val)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "py39",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.13"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}