File size: 6,565 Bytes
5effdd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
import chex
import jax
from jax import random, jit, vmap
import jax.numpy as jnp
from functools import partial
from consts import EPS
from globals import UserInfo, Char, State


def norm_playtime(arr: chex.Array, cid: int) -> chex.Array:
    max_playtime = jnp.max(arr) + EPS
    norm = arr[cid] / max_playtime
    return norm


@jit
def construct_feats(user: UserInfo, char: Char, char_id: int) -> chex.Array:
    feats = [
        user.skill_level,
        jnp.log1p(user.games_played),
        char.difficulty,
        char.execution_level,
        char.neutral_required,
        char.tier,
    ]
    feats.append(char.archetype_vec)
    skill_match = 1.0 - jnp.abs(user.skill_level - (1.0 - char.difficulty))

    feats.append(jnp.array([skill_match]))

    archetype_sim = jnp.dot(user.pref_archetype, char.archetype_vec)
    feats.append(jnp.array([archetype_sim]))

    tried_before = user.chars_attempted_mask[char_id]
    novelty_bonus = 1.0 - tried_before
    feats.append(jnp.array([novelty_bonus]))

    past_perf = jnp.where(tried_before > 0.5, user.wr[char_id], 0.5)

    feats.append(jnp.array([past_perf]))

    norm = norm_playtime(user.playtime, char_id)
    feats.append(jnp.array([norm]))

    return jnp.concatenate([jnp.atleast_1d(feat) for feat in feats])


@partial(jit, static_argnums=(2,))
def build_feats(user: UserInfo, chars: Char, n_chars: int):
    def build_single(cid: int):
        char = jax.tree.map(lambda x: x[cid], chars)
        return construct_feats(user, char, cid)

    return vmap(build_single)(jnp.arange(n_chars))


@jit
def sample_params(key: chex.PRNGKey, mu: chex.Array, Sigma: chex.Array) -> chex.Array:
    d = mu.shape[0]
    Lambda = Sigma + EPS * jnp.eye(d)
    theta = random.multivariate_normal(key, mu, Lambda)
    return theta


@jit
def compute_expected_rewards(thetas: chex.Array, feats: chex.Array) -> chex.Array:
    return vmap(jnp.dot)(thetas, feats)


@jit
def thompson_sample(
    key: chex.PRNGKey, state: State, feats: chex.Array
) -> tuple[chex.Array, chex.Array]:
    num_chars = feats.shape[0]
    keys = random.split(key, num_chars)

    thetas = vmap(sample_params)(keys, state.mu, state.Sigma)
    rewards = compute_expected_rewards(thetas, feats)
    return rewards, thetas


@jit
def update_posterior(
    state: State,
    char_id: int,
    feats: chex.Array,
    reward: float,
    noise_var: float = 1.0,
    use_adaptive_noise: bool = True,
) -> State:
    x = feats
    d = x.shape[0]
    mu_old = state.mu[char_id]
    sigma_old = state.Sigma[char_id]

    # might be numerically unstable, not sure... for noninvertivle matrices should check this later when not lazy
    # ugly and hacky but idk how to approx this outside of inv, solve and do op, then inv to undo

    Sigma_old_inv = jnp.linalg.inv(sigma_old + EPS * jnp.eye(d))
    Sigma_new_inv = Sigma_old_inv + (1.0 / noise_var) * jnp.outer(x, x)
    Sigma_new = jnp.linalg.inv(Sigma_new_inv)

    mu_new = Sigma_new @ (Sigma_old_inv @ mu_old + (reward / noise_var) * x)

    new_mu = state.mu.at[char_id].set(mu_new)
    new_Sigma = state.Sigma.at[char_id].set(Sigma_new)

    # TODO: figure out whether adaptive noise in gp is needed
    new_beta = None

    if use_adaptive_noise:
        new_beta = state.beta.at[char_id].add(1)
    return State(
        mu=new_mu,
        Sigma=new_Sigma,
        alpha=state.alpha,
        beta=new_beta if new_beta is not None else state.beta,
    )


@partial(jit, static_argnums=(2, 3))
def select_top_k_diverse(
    scores: chex.Array, archetypes: chex.Array, k: int, diversity_thresh: float
) -> chex.Array:
    n_chars = scores.shape[0]
    sorted_idx = jnp.argsort(-scores)

    def selection_step(carry, cand_idx):
        select, cnt = carry
        cand_idx = sorted_idx[cand_idx]

        done = cnt > k

        cand_arch = archetypes[cand_idx]

        def check_item_diversity(sel_idx):
            # may need a max bound here
            is_valid = sel_idx >= 0
            sel_arch = archetypes[sel_idx]
            # cos_sim w little eps to avoid div 0

            sim = jnp.dot(cand_arch, sel_arch) / (
                jnp.linalg.norm(cand_arch) * jnp.linalg.norm(sel_arch) + 1e-8
            )
            return jnp.where(is_valid, sim < diversity_thresh, True)

        all_diverse = jnp.all(vmap(check_item_diversity)(select))

        add_op = jnp.logical_and(jnp.logical_not(done), all_diverse)

        new_sel = jnp.where(add_op, select.at[cnt].set(cand_idx), select)
        new_cnt = jnp.where(add_op, cnt + 1, cnt)
        return (new_sel, new_cnt), None

    init = jnp.full(k, -1, dtype=jnp.int32)
    init = init.at[0].set(sorted_idx[0])

    (final_sel, null), null = jax.lax.scan(
        selection_step, (init, 1), jnp.arange(1, n_chars)
    )
    return final_sel


@jit
def compute_reward(
        won: bool, completed: bool, rating: float, playtime_mins: float, weights:chex.Array = jnp.array([0.3, 0.15, 0.25, 0.3])
) -> float:
    win_reward = jnp.where(won, weights[0], 0.0)
    completion_reward = jnp.where(completed, weights[1], 0.0)
    rating_reward = weights[2] * jnp.clip(rating / 5.0, 0.0, 1.0)

    engagement_reward = weights[3] * jnp.clip(jnp.log1p(playtime_mins) / 5.0, 0.0, 1.0)
    return win_reward + completion_reward + rating_reward + engagement_reward


@partial(jit, static_argnums=(4, 5))
def recommend_characters(
    key: chex.PRNGKey,
    state: State,
    user: UserInfo,
    characters: Char,
    n_chars: int,
    top_k: int = 3,
    diversity_threshold: float = 0.75,
) -> tuple[chex.Array, chex.Array]:
    features = build_feats(user, characters, n_chars)
    sampled_rewards, sampled_thetas = thompson_sample(key, state, features)

    selected = select_top_k_diverse(
        sampled_rewards, characters.archetype_vec, top_k, diversity_threshold
    )

    return selected, sampled_rewards


def init_thompson(n_chars: int, feature_dim: int, prior_var: float = 1.0) -> State:
    return State(
        mu=jnp.zeros((n_chars, feature_dim)),
        Sigma=jnp.tile(prior_var * jnp.eye(feature_dim), (n_chars, 1, 1)),
        alpha=jnp.ones(n_chars),
        beta=jnp.ones(n_chars),
    )


@jit
def batch_update_posterior(
    state: State,
    char_ids: chex.Array,
    features: chex.Array,
    rewards: chex.Array,
    noise_var: float = 1.0,
) -> State:
    def single_update(s, data):
        char_id, feat, reward = data
        return update_posterior(s, char_id, feat, reward, noise_var), None

    final_state, _ = jax.lax.scan(single_update, state, (char_ids, features, rewards))
    return final_state