| 89 | |
| 90 | |
| 91 | def evolution_strategy( |
| 92 | f, |
| 93 | population_size, |
| 94 | sigma, |
| 95 | lr, |
| 96 | initial_params, |
| 97 | num_iters): |
| 98 | |
| 99 | # assume initial params is a 1-D array |
| 100 | num_params = len(initial_params) |
| 101 | reward_per_iteration = np.zeros(num_iters) |
| 102 | |
| 103 | params = initial_params |
| 104 | for t in range(num_iters): |
| 105 | t0 = datetime.now() |
| 106 | N = np.random.randn(population_size, num_params) |
| 107 | |
| 108 | # ### slow way |
| 109 | # R = np.zeros(population_size) # stores the reward |
| 110 | |
| 111 | # # loop through each "offspring" |
| 112 | # for j in range(population_size): |
| 113 | # params_try = params + sigma*N[j] |
| 114 | # R[j] = f(params_try) |
| 115 | |
| 116 | ### fast way |
| 117 | R = pool.map(f, [params + sigma*N[j] for j in range(population_size)]) |
| 118 | R = np.array(R) |
| 119 | |
| 120 | m = R.mean() |
| 121 | A = (R - m) / R.std() |
| 122 | reward_per_iteration[t] = m |
| 123 | params = params + lr/(population_size*sigma) * np.dot(N.T, A) |
| 124 | print("Iter:", t, "Avg Reward:", m, "Duration:", (datetime.now() - t0)) |
| 125 | |
| 126 | return params, reward_per_iteration |
| 127 | |
| 128 | |
| 129 | def reward_function(params): |