(self)
| 899 | self._init_params() |
| 900 | |
| 901 | def _init_params(self): |
| 902 | E = self.env_info |
| 903 | assert not E["continuous_actions"], "Action space must be discrete" |
| 904 | |
| 905 | obs_encoder = None |
| 906 | if E["continuous_observations"]: |
| 907 | obs_encoder, _ = tile_state_space( |
| 908 | self.env, |
| 909 | self.env_info, |
| 910 | self.n_tilings, |
| 911 | state_action=False, |
| 912 | obs_max=self.obs_max, |
| 913 | obs_min=self.obs_min, |
| 914 | grid_size=self.grid_dims, |
| 915 | ) |
| 916 | |
| 917 | self._create_2num_dicts(obs_encoder=obs_encoder) |
| 918 | |
| 919 | # behavior policy is stochastic, epsilon-soft policy |
| 920 | self.behavior_policy = self.target_policy = self._epsilon_soft_policy |
| 921 | if self.off_policy: |
| 922 | # target policy is deterministic, greedy policy |
| 923 | self.target_policy = self._greedy |
| 924 | |
| 925 | # initialize Q function |
| 926 | self.parameters["Q"] = defaultdict(np.random.rand) |
| 927 | |
| 928 | # initialize returns object for each state-action pair |
| 929 | self.derived_variables = {"episode_num": 0} |
| 930 | |
| 931 | self.hyperparameters = { |
| 932 | "agent": "TemporalDifferenceAgent", |
| 933 | "lr": self.lr, |
| 934 | "obs_max": self.obs_max, |
| 935 | "obs_min": self.obs_min, |
| 936 | "epsilon": self.epsilon, |
| 937 | "n_tilings": self.n_tilings, |
| 938 | "grid_dims": self.grid_dims, |
| 939 | "off_policy": self.off_policy, |
| 940 | "temporal_discount": self.temporal_discount, |
| 941 | } |
| 942 | |
| 943 | self.episode_history = {"state_actions": [], "rewards": []} |
| 944 | |
| 945 | def run_episode(self, max_steps, render=False): |
| 946 | """ |
no test coverage detected