| 246 | |
| 247 | |
| 248 | ActionPtr ModelReusableAgent::selectNewAction() { |
| 249 | ActionPtr action = nullptr; |
| 250 | action = this->selectUnperformedActionNotInReuseModel(); |
| 251 | if (nullptr != action) { |
| 252 | BLOG("%s", "select action not in reuse model"); |
| 253 | return action; |
| 254 | } |
| 255 | |
| 256 | action = this->selectUnperformedActionInReuseModel(); |
| 257 | if (nullptr != action) { |
| 258 | BLOG("%s", "select action in reuse model"); |
| 259 | return action; |
| 260 | } |
| 261 | |
| 262 | action = this->_newState->randomPickUnvisitedAction(); |
| 263 | if (nullptr != action) { |
| 264 | BLOG("%s", "select action in unvisited action"); |
| 265 | return action; |
| 266 | } |
| 267 | |
| 268 | // if all the actions are explored, use those two methods to generate new action based on q value. |
| 269 | // there are two methods to choose from. |
| 270 | // based on q value and a uniform distribution, select an action with the highest value. |
| 271 | action = this->selectActionByQValue(); |
| 272 | if (nullptr != action) { |
| 273 | BLOG("%s", "select action by qvalue"); |
| 274 | return action; |
| 275 | } |
| 276 | |
| 277 | // use the traditional epsilon greedy strategy to choose the next action. |
| 278 | action = this->selectNewActionEpsilonGreedyRandomly(); |
| 279 | if (nullptr != action) { |
| 280 | BLOG("%s", "select action by EpsilonGreedyRandom"); |
| 281 | return action; |
| 282 | } |
| 283 | BLOGE("null action happened , handle null action"); |
| 284 | return handleNullAction(); |
| 285 | } |
| 286 | |
| 287 | /// Randomly choose an action that is not been visited before, which belongs to the following type: |
| 288 | /// BACK, |
no test coverage detected