An architect performing offline planning and real-time NLG.

The behaviour is configured with a config file; an example is the architect-config.yaml file, which is also read by default.

You can use a different config file by supplying it as the first argument. You can additionally modify the port via the second argument (which takes precedence over the port configured in the config file).

running the architect server

With gradle, use this to run the architect server:

./gradlew run --args="my-config.yaml"

configuration

There are currently four different architects:

• BLOCK: instructs block by block
• MEDIUM: teaches each high-level concept once block by block and then uses them
• HIGHLEVEL: Only instructs high-level objects (wall, floor, railing, ...)
• adaptive: simulates all three above and picks the one with the lowest overall cost according to the generation model (i.e. sum of the cost of each instuction)

you can swith between these by setting the instructionlevel.

The Architect uses minecraft-nlg as the sentence generator. minecraft-nlg is based on a grammar (see Köhn and Koller, 2019) and computes the optimal instruction based on the grammar's weights. By default, these are nearly all the same, i.e. it always generates the shortest instruction. The weights can also be learned from logs previous games, by using the weight-estimator. For this, one has to set

• useTrainedWeights: true
• weightTrainingDatabase: "jdbc:mariadb://localhost:3306/SOMEDATABASE" (where the database string is the one where the games are stored)

Then the weight estimator will run bootstrapping to generate sets of grammar weights, read out the lower and higher percentile you configured (trainingSamplingLowerPercentile and trainingSamplingUpperPercentile) and set each weight to a random sample from the uniform distribution between the lower and upper percentile. This is by design not the optimal set of weights as it still enables exploration for further games. The weights are resampled for every new game.

Alternatively, one can use random weights. Then the weights will be randomized for every game on a uniform scale between 1 and 10. Set randomizeWeights: true for this.

When running experiments, you can set a secret word to be shown once the requirements to get paid are fulfilled. We use this together with an external form which only continues after entering the secret word. set secretWord: MYSECRETWORD. This word is shown when the instructions were completed or alternatively if a number of correct blocks has been placed and a certain amount of time has passed. You can configure both with timeoutMinutes: 10 and timeoutMinBlocks: 4.

The configs/ directory contains a selection of different configurations.