def run_dqn(name_asset, n_features, n_neurons, n_episodes, batch_size,
random_action_decay, future_reward_importance):
# returns a list of stocks closing price
df = pd.read_csv(INPUT_CSV_TEMPLATE % name_asset)
data = df['Close'].astype(
float).tolist() #https://www.kaggle.com/camnugent/sandp500
l = len(data) - 1
print(
f'Running {n_episodes} episodes, on {name_asset} (has {l} rows), features={n_features}, '
f'batch={batch_size}, random_action_decay={random_action_decay}')
dqn = Dqn()
profit_vs_episode, trades_vs_episode, epsilon_vs_episode, model_name, num_trains, eps = \
dqn.learn(data, n_episodes, n_features, batch_size, USE_EXISTING_MODEL, RANDOM_ACTION_MIN,
random_action_decay, n_neurons, future_reward_importance)
print(f'Learning completed. Backtest the model {model_name} on any stock')
print('python backtest.py ')
print(f'see plot of profit_vs_episode = {profit_vs_episode[:10]}')
plot_barchart(profit_vs_episode, "episode vs profit", "episode vs profit",
"total profit", "episode", 'green')
print(f'see plot of trades_vs_episode = {trades_vs_episode[:10]}')
plot_barchart(trades_vs_episode, "episode vs trades", "episode vs trades",
"total trades", "episode", 'blue')
text = f'{name_asset} ({l}), features={n_features}, nn={n_neurons},batch={batch_size}, ' \
f'epi={n_episodes}({num_trains}), eps={np.round(eps, 1)}({np.round(random_action_decay, 5)})'
print(f'see plot of epsilon_vs_episode = {epsilon_vs_episode[:10]}')
plot_barchart(epsilon_vs_episode, "episode vs epsilon",
"episode vs epsilon",
"epsilon(probability of random action)", text, 'red')
print(text)
future_reward_importance = 0.9500 # (float) 0-1 aka decay or discount rate, determines the importance of future
# rewards.If=0 then agent will only learn to consider current rewards. if=1 it will make it strive for a long-term
# high reward.
# do not touch those params
random_action_min = 0.0 # (float) 0-1 do not touch this
use_existing_model = False # (bool) do not touch this
data = getStockDataVec(stock_name) # https://www.kaggle.com/camnugent/sandp500
l = len(data) - 1
print(
f'Running {episodes} episodes, on {stock_name} (has {l} rows), features={num_features}, batch={batch_size}, random_action_decay={random_action_decay}'
)
dqn = Dqn()
profit_vs_episode, trades_vs_episode, epsilon_vs_episode, model_name, num_trains, eps = \
dqn.learn(data, episodes, num_features, batch_size, use_existing_model, random_action_min, random_action_decay,
num_neurons, future_reward_importance)
print(
f'i think i learned to trade. now u can backtest the model {model_name} on any stock'
)
print('python backtest.py ')
minutes = np.round((time.time() - start_time) / 60, 1) # minutes
text = f'{stock_name} ({l}),t={minutes}, features={num_features}, nn={num_neurons},batch={batch_size}, epi={episodes}({num_trains}), eps={np.round(eps, 1)}({np.round(random_action_decay, 5)})'
print(f'see plot of profit_vs_episode = {profit_vs_episode[:10]}')
plot_barchart(profit_vs_episode, "episode vs profit", "episode vs profit",
"total profit", "episode", 'green')
print(f'see plot of trades_vs_episode = {trades_vs_episode[:10]}')
plot_barchart(trades_vs_episode, "episode vs trades", "episode vs trades",
"total trades", "episode", 'blue')
state.next_state = State(next_state)
if (learn):
nnet.addToMemory(state) # memorize the step transition values
state = state.next_state
episode_rewards += reward
if done:
RESULT_TOTALS.append(episode_rewards)
break
if (learn):
nnet.learn(np.average(RESULT_TOTALS))
else:
if (LEARN_SAVE):
save(nnet)
LEARN_SAVE = False
plt.plot(RESULT_TOTALS)
plt.grid(b=True,
which='major',
axis='y',
color='r',
linestyle='-',
linewidth=.5)
plt.show()
print(RESULT_TOTALS)