Development Showcase: AI-Controlled Companies
Hola, it's the developer of Harwell - Caleb - here! In this development showcase, I'll be covering the implementation of AI-controlled players and how they'll interact with the systems and compete with the player.
This Dev diary is not about LLMs and stuff, just about the game "AI".
Names & Characteristics
The AI get given randomly selected names based on historical space-related scientists. Not just from Europe or the Americas, but from Asia, Africa, and the rest of the world also. The list of names will eventually be a moddable attribute so players can add their own names.
The AI are also treated the same as players - given a player colour, company values and stock prices, and everything else a player gets. The game considers them a player, just with AI-specific functions in certain areas (an AI-specific market purchase, AI-specific construction) for better compatibility with their inventory.
Logic
Weight-Based Decision Making
The challenge of making a solid AI that's both competitive but not impossible to beat, yet fun to play against, is a tricky one. In order to make it as realistic and human as possible, I've opted to use a weight based decision making system.
The logic process would go like this:
Weigh up all decisions in terms of various attributes and rank the best options.
Assign a weighted value to each position, so there's a high chance of picking the best option, then a slightly lower chance of picking the next best, and so on...
Randomly make a choice, taking the weighted values into consideration.
This should keep the reasoning behind the decisions whilst adding some randomness. It also means that the AI doesn't make the perfect choice every time. Another benefit of this system, is that I can export those weights and values and easily adjust them in order to balance the difficulty of the AI. These values can be exported as part of the modding system so players can create custom AI logic, though I'll worry about the exact modding implementation afterwards.
Attributes
The AI 'brain' is composed of multiple attributes, these are the attributes I'll eventually open up to the game's modding interface so that it can be adjusted later on! So far, the AI has various values:
Risk Tolerance Adjusts the cash reserve ratios and spending decisions based on risk.
Check Interval How often to make general decisions.
Building Check Interval How often to make building decisions.
Cash Reserve Ratio The minimum cash an AI must have before it can make spending decisions.
Max Inventory Ratio How many goods an AI can have maximum before they should be selling.
Reserve Stock Ratio The minimum amount of goods the AI should keep in reserve.
Stock Budget Ratio The ratio of an AI's money they will spend purchasing stock in other players.
Sell thresholds: How aggressive the AI should be with selling their inventory. The lower the threshold the more aggressive the selling.
Sell Threshold - High
Sell Threshold - Medium
Sell Threshold - Emergency
There's also Building Weight Values, which is the effect attributes have on an AI's decision making:
Profitability The most important factor when it comes to an AI decision.
Market Demand The AI also considers how much demand there is for a good.
Supply Chain The AI aims to build up a supply chain of goods. For example, if they have a Steel Mill, an Iron Mine will have a high supply chain value.
Risk Factor The risk associated with a purchasing decision. How much will the AI have in cash reserves if it goes ahead with this?
Spawning Logic
The AI spawning Logic is kept quite simple for the moment, they place their HQs on a somewhat random tile on the map. However, they consider two things before they make their decision:
Distance to resources - the AI wants to be somewhere they can harvest resources easily. Distance doesn't currently have an effect in the game, but buildings will soon cost more the further away they are built from your HQ.
Distance from other players - the AI will try and avoid other players to prevent crowding and overlap. This isn't important it just makes the map a little nicer and stops them from griefing players.
Market Logic
The AI won't often make market purchasing decisions, as they will primarily focus on constructing buildings and those buildings will manage the AI 'purchasing' of goods. Sometimes though, the AI might invest in a certain good if it thinks it will get a return. Currently I've mostly disabled purchasing as it's not properly implemented, but this will be improved in future updates.
The goal is to have the AI try to predict future actions - buy cheap, sell high. They'll buy Steel if the price is well below the market base price, especially if they think the price is going to go up. They can then sell those goods later for a profit.
Building Logic
An AI makes a multi-step construction decision similar to how it builds it's HQ. It takes into consideration the AI attributes we covered at the start.
The logic goes as follows:
Work out the value of each potential building based on profitability, market demand, supply chain, and risk factor.
Select the one of the best options.
Check if the building is affordable by the AI based on risk tolerance and cash reserve values.
Purchase the building and place it on the map. Placement is currently random (unless it's something like a mine).
Stock Logic
Stock logic is rudimentary at the moment, and it does purchase player and AI stocks, but the stock menu is still under development. The goal is to have the AI evaluate the players in the game, work out who best to buy stocks in and whether it is worth it, and then purchase those stocks. They'll weigh up buying multiple stocks in one player vs diversifying their portfolio into multiple players.
Behaviour
Personalities
One feature that will be coming to AI in the future is personalities. Those attributes mentioned before? When those are abstracted into JSON files for modding, the game will have various personality files and will allow you to add your own. The AI will then be given a personality file which defines how they play.
Using this, we can create different AI types that play vastly differently from each other:
Aggressive spenders: AI that has a high risk tolerance and minimum cash reserves. Makes more risky plays which could lead to more profit, but opens it up to certain vulnerability.
Tame spenders: Low risk tolerance, low cash reserves. They'll spend money, but they take more care in each decision and won't make purchases they aren't confident in.
Aggressive hoarders: High risk tolerance but higher cash reserves, they're more resilient to changes in the market, but when they're willing to take risks as well.
etc.
Of course these will take some fine tuning and attributes may be added/changed as time goes on. There will be a lot to balance in the game and this is one of them, but the way the AI is built means these values can be changed on the fly and tested straight away without rewriting large parts of the code. It also means we can create difficulty levels based on how well certain personalities do.
Difficulty
I don't want difficulty to be 'the AI gets flat bonuses' or 'the player starts with a handicap'. I don't like those forms of difficult levelling and I don't think they're realistic or fun to play with.
Depending on the outcome of AI testing, I want difficult to be made up of AI personalities. Perhaps a higher difficulty level introduces more aggressive AI, whilst lower difficulty levels utilises the tame AI or the AI with higher cash reserves. There's no flat modifiers for you to take advantage of, the AI is instead like playing against a 'good player' vs a 'bad player' at the game. Difficulty will take some balancing as well, and I'll write a future dev diary on it as it isn't introduced yet, but that's the plan at least!
Winning
Of course, the AI will strive to win. It might not be the best at the game, but they're no chump either. Each AI will be told the goal of the game, but it's ultimately down to the attributes defined as to how well they do at actually winning the game. This will take some balancing and adjustment, and most importantly requires me to get the stock menu sorted, so stick around for that dev diary!
Anyway, that's all from me on this topic, I hope you liked it! I think this one is one of the most interesting development diaries so far, and I'm excited to work on and improve the AI further.
Ciao,
Caleb